• BJOMATHEMATICS TRAINING PROGRAM The Locally Optimal Combination of Certain Multivariate Test Statistics by Katherine Lane Monti May 15, 1975 Institute of Statistics Mimeo Series /I 1(;) t:J" MONTI, KATHERINE LANE. The Locally Optimal Combination of Certain Multivariate Test Statistics. (Under the direction of PRANAB KUMAR SEN and DANA QUADE.) Most overall tests of hypotheses combining s ~ 2 independent statistics are based on the p-values of the individual tests. Van Zwet and Oosterhoff considered procedures for combining tests of univariate hypotheses which are based on the values of the (asymptotically exponentially distributed) statistics themselves. This dissertation deals with combinations of certain multivariate test statistics with distributions which are not in the expone~tial family • . Many of the classical GLMM analysis of variance statistics have F distributions and can be transformed tIJ beta-distributed random variables. It is shown that if the statistics to be combined have beta distributions which are noncentral under the alternative hypothesis, then a type D test, which provides a locally most powerful combination procedure for testing the null hypothesis does exist. The form of the test is found for the balanced case where the design matrices are proportiona1. For the unbalanced case, a test which has the best average local power is found; this test is termed a type F test. It is shown that, for local alternatives, if there are a large number of small-sample statistics to be combined, the type D comb ination procedure is more efficient than Fisher's method according to Bahadur asymptotic relative efficiency. It is also shown that under the same assumptions, the type F combination procedure is more efficient than Fisher's according to an average measure of Bahadur asymptotic relative efficiency. The exact distribu.tion of the overall type D and type F statistics is given if s = 2. As s increases, the exact distribution of the statistics becomes more difficult to calculate. The Central Limit Theorem does apply, however, and the accuracy of applying the Central Limit Theorem to find approximate critical values for the overall test statistic is discussed. The method of applying the type D and type F tests is illustrated in a numerical example based on some pharmaceutical data. THE LOCALLY OPTIMAL COMBINATION OF CERTAIN MULTIVARIATE TEST STATISTICS by Katherine Lane Monti A Dissertation submitted to the faculty of the University of North Carolina at Chapel Hill in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Biostatistics. Chapel Hill 1975 Approved by I f ACKNOWLEDGMENTS The author appreciates the assistance of the members of her advisory committee, Drs. J. C. Cornoni, J. E. Grizzle, R. R. Kuebler, D. Quade, and P. K. Sen. Special thanks go to Dr. P. K. Sen for his guidance of this research and his reassuring support throughout the duration of this undertaking, and to Dr. D. Quade, who adopted the responsibilities of adviser when Dr. Sen was on sabbatical, for his patient assistance. Thanks are also due to Drs. T•. Donne11y and J. E. Grizzle, who were responsible for the author's initial interest in biostatistics. The financial assistance of the Department of Biostatistics during the course of the author's graduate studies is gratefully acknowledged. This research was supported by National Institues of Health, Division of General Medical Sciences Grant Nos. 5TDl GM00038-20 and 5T01 GM0003~-21. The author wishes to thank her husband, Danny, for his continued support and encouragement throughout the course of the author's studies and especially for his understanding while the author was preparing for and taking the Basic Examination. For her typing of the manuscript, Mrs. Gay Hinnant is. warmly thanked. ii TABLE OF CONTENTS Page ACKNOWLEDGMENTS. .... ii LIST OF TABLES • • v LIST OF FIGURES. vi Chapter I. INTRODUCTION. 1 1.1. Statement of the Problem and Outline of the Research. • • • • • • • • • 1.2. Review of the Literature •••• 1.3. Introduction to Type D and Type F Critical Regions • • • 1 4 9 1.3.1. Type D Critical Regions • 1.3.2. Type F Critical Regions •• 9 14 II. TYPE D CRITICAL REGIONS FOR COMBINING BETADISTRIBUTED STATISTICS •• 17 17 17 2.1. Introduction • • • • 2.2. Permissible Tests and Critical Regions 2.3. The Existence of a Type D Critical Region for Combining s Independent Statistics with Beta Distributions • . • • 2.4. The Type D Critical Region: Balanced Design. • • • • • • • • • • • • • 2.5. The Type D Critical Region for a Size-a Combination Procedure in the Two-Sample Case 20 26 28 29 35 37 2.5.1. Case 1. 2.5.2. Case 2•• 2.5.3. Case 3. 2.6. The s-Sample Case • • • • • • 2.7. A Note on Applications. iii \ \, .... 39 40 Page III. FURTHER CONSIDERATIONS IN THE LOCALLY OPTIMAL COMBINATION OF BETA-DISTRIBUTEp STATISTICS • 3.1. Introduction. • • • 3.2. The Type F Critical Region: Unbalanced Design • 3.3. Bahadur Efficiency. 3.4. Extensions to Other Problems. IV. APPROXIMATE CRITICAL VALUES FOR TYPE D AND TYPE F TESTS • • • • • • • • • • • 42 42 42 46 55 60 V. TABLES OF CRITICAL VALUES FOR TYPE D AND TYPE F TESTS AND A NUMERICAL ILLUSTRATION • • • • • • • • • 65 5.1. Tables of Critical Values for Type D and Type F Tests 5.2. Numerical Illustration. 80 VI. SUMMARY AND SOME SUGGESTIONS FOR FURTHER RESEARCH • • • • 65 86 APPENDIX A: USEFUL MATHEMATICAL RESULTS 88 BIBLIOGRAPHY .............. 95 iv LIST OF TABLES Table Page 4.1. Values of the Kolmogorov-Smirnov Statistic for Certain Simulation Problems, Assuming Certain Parameters are Constant for All Samples • • • . 63 4.2. Values of the Kolmogorov-Smirnov Statistic for Certain Simulation Problems Without Restriction on the Constancy of Parameters • • • • • • • • • 64 5.1.1. Critical Values for Two-Sample Type D and Type F Tests. • 5.1.2. Estimated Critical Values for Type D and Type F Tests for Certain Interpolated and Extrapolated Values of the Parameters a , a , and f . 2 1 5.2.1. Distribution of Patients Among Six Hypertension Treatments in Two Clinics. v \ 66 79 80 LIST OF FIGURES Page Figure 2.5.1. Type D Critical Regions for Combining Two Beta-Distributed Statistics • • • • • 29 3.2.1. The Ellipse Formed by the Approximate Intersection of the Power Surface and a Horizontal Plane at a Height of a + E • 45 vi CHAPTER I INTRODUCTION 1.1. Statement of the Problem and Outline of the Research Let Ql, ••• ,Qs be s independent statistics for testing the same null hypothesis against the same alternative. The problem is to combine the results of these s tests into a single overall test for the common testing problem. Most combination procedures available in the litera- ture are based not on the statistics Q. but on their probability inte1. gral transforms Pi' where Pi value of Q for i i = Pr(Q.1 .>- = l, ••• ,s. q.1. IH0 ) whereq.1. is the observed Those combination procedures which have been based on the Q. 's are mainly limited to the case of univariate 1. hypotheses with test statistics distributed in the one-parameter exponential family. The aim of this dissertation is to extend these latter results to certain multivariate and multiparametric testing problems. It is assumed throughout that the distributions of the statistics are continuous. Let the hypothesis be H: o A = _0 A (1.1.1) with alternative A ., A - -0 Statistics used in the testing of this hypothesis include Hotelling's 2 T statistics and statistics used in conjunction with the general linear j 2 model. If X = ~§ + £, where ~ NXq design matrix, ~ is an NXm matrix of observations, A is an is a qXm matrix of unknown parameters, and £ is an NXm error matrix, then the general linear hypothesis is R: o C~M = 0 with alternative where ~ and of zeros. ~ are appropriate gXq and mxu matrices and 0 is a gXu matrix If g=p and u=l or g=l and u=p, then tor, respectively. specified above. C~M is a pXl or lXp vec- In either case, the null hypothesis is of the form Now assume that C and M are of full rank and let - -- - - - -- E = M' X' [I-A(A' A) -lA'] XM. - - - - -- R = M'X'A(A'A)-lC' [C(A'A)-\;;'J-lC(A'A)-lA'XM and - - - --- Statistics commonly used in testing (1.1.1) are based on the characteristic roots t , i = l, ••• ,[min(g,u)] of Ig-t~l: Roy's largest root i 2 (max t.); the Rotelling-Lawley T (~t.); and Wilks' ~ (TI(l+t.)-l). In i 1 i 0 1 i 1 the current problem there is only one root, t, so that Roy's statistic 2 and T are identical. o Under the null hypothesis and assumptions of normalitY,(N-q)t/g = F is an F-distributed variate with g and N-q degrees of freedom. The above model can, for the type of hypothesis being considered, be restated as R : o where a is a non-null vector. - C~Ma ---- :f 0, Then the common test statistic is F(a) = (N-q)/g[a'Ra/a'Ea], which has an F distribution with g and N-q - -- - degrees of freedom. -- (Anderson [1958], Morrison [1967]). Since the 'F distribution is difficult to work with, it is assumed that the statistics have been transformed to beta-distributed random variables. If F has an F distribution with p and q degrees of 3 freedom, then U p/2 and q/2. = (pF/q)/(l+pF/q) has a beta distribution with parameters If F is noncentrally distributed with parameter 8 t U has a noncentral beta distribution with the same noncentrality parameter. If F has a noncentral F distribution with noncentrality parameter 8 and degrees of freedom p and q, then the density of F is given by -8/2 ~ 00 1: r=O q 8r (pf/q)p/2+r-l 2r rlS(p/2+r t q/2)(1+pf/q)(P+q)/2+r ' 0, f > 0; f < o. When 8=0, the density reduces to that of a central F distribution, given by (p/q)p/2 f P/ 2- l /[S(p/2,q/2) for positive values of f. (l+pf/q) (P+q)/2] If F has a noncentral distribution and U is as defined above, then U has a noncentral beta distribution with noncentrality parameter 8 and parameters p/2 and q/2. The density of U is given by 00 e- 8 / 2 E o 2. u 2 0, otherwise. r=O 1; (1.1. 2) When 8=0, U has a central beta distribution and density (1.1.3) O<u<l. In this dissertation some locally most powerful combinations of beta-distributed statistics used in tests of (1.1.1) are developed. The concepts and notions of type D and type F tests, which maximize different measures of average curvature of the power surface at A , are .... 0 4 presented in Section 1.3. In Chapter II the type D test is determined for designs which are balanced as defined in Section 2.3, with details for the two-sample case. In Chapter III the type F test, which can be applied in the case of an unbalanced design, is determined. It is also shown that if there are a large number of finite-sample statistics to be combined, the test based on the type D region is more efficient than Fisher's test when the comparison is made on the basis of Bahadur asymptotic relative efficiency. Furthermore, it is shown that the type F test is more efficient than Fisher's test when the comparison is made on the basis of a measure of average Bahadur asymptotic relative efficiency. A sec- tion mentioning some extensions concludes this chapter. The two-sample results discussed in Chapter II are difficult to extend to cases involving even a moderately large number of statistics to be combined. However, the overall statistic is a linear combination of the independent statistics to be combined, and the Central Limit Theorem does apply. Finite-sample approximations to this asymptotic result are studied in Chapter IV. Chapter V includes an example of applications of the theory to some clinical data. Some tables of critical values for the two-sample case are also presented. A final summary chapter concludes the disserta- tion. 1.2. Review of the Literature Recall that the problem is to combine the results of s independent tests into a single overall test for the common testing problem. 5 Let Qi' i = l, ••• ,s, be the test statistics and let Pi' i = l, ••• ,s, be their probability integral transforms, as defined in the preceding section. Many of the combination procedures in the literature are based only on the Pi's. x= Fisher [1932, pp. 99-101] proposed -2 log Pl··· Ps as an overall test statistic. Under the null hypothesis, X is distributed as a chi-squared random variable with 2s degrees of freedom. Independently of Fisher, K. Pearson [1933] proposed as an appropriate statistic. Further discussion of these methods is given by E. S. Pearson [1938], Wallis [1942] and Oosterhoff [1969]. Tippett [1931] proposed that the null hypothesis be rejected if the smallest of the P. 's is sufficiently small. 1. Wilkinson [1951] later extended this test, suggesting that the k-th smallest of the P. 's 1. be used as the statistic for the overall test, where k is a predetermined value. Liptak [1958] proposed combinations of the form where -1 ~ is the inverse of an arbitrary function ¢ and the w. are suitable weights. 1. Liptak suggested that one natural choice of ¢ is the normal distribution function. function, so that ¢ -1 If ~ is taken to be the exponential is the logarithmic function, this test reduces to a weighted version of Fisher's test, proposed earlier by Good [1955] and applied by Zelen and Joel [1959] in the case where the Q. are distrib1. uted according to an F distribution and s = 2. Lancaster [1961] also 6 discussed adding weights to Fisher's method. Birnbaum [1954] compared the methods of Fisher, Pearson, and Wilkinson. lems. He found that no method is optimal for all combination prob- However, he showed that if the test statistics have distributions belonging to the exponential class, then Pearson's method and Wilkinson's method with k>2 are inadmissible. discussed in Section 2.2.) (Admissibility is defined and He also noted that while Tippett's method is admissible, Fisher's method seems to be somewhat more uniformly sensitive to commonly chosen alternatives. Naik [1969] compared Fisher's method to Tippett's method, and Wilk and Shapiro [1968] compared Fisher's method to the mean of the normal transform for empirical studies. In both papers it was noted that Fisher's method is not necessarily optimal for all testing problems. Littell and Folks [1971] compared Fisher's method, Tippett's method, Wilkinson's method with k=s, and Liptak's method, with ¢ defined as the normal distribution function, using Bahadur relative efficiency. Assuming that there were a fixed number of large-sample statistics to be combined, the authors showed that Fisher's method is generally at least as efficient as the other methods. However, for some testing problems, not all methods can be distinguished on the basis of Bahadur efficiency and in these cases other methods may be superior. In a later paper, Littell and Folks [1973] showed that, under the same assumption, Fisher's method is at least as good as any other method on the basis of Bahadur relative efficiency. The problem of finding optimal combination procedures is still open, however, since it is often the case that researchers have sample 7 sizes insufficient to satisfy the asymptotic arguments of Bahadur efficiency. Furthermore, as noted by Littell and Folks, Bahadur effi- ciency cannot distinguish Fisher's method from some other methods for certain testing problems. All of the tests discussed thus far are based on the probability integral transforms of the s test statistics which are to be combined. Various justifications for using the transforms rather than the test statistics themselves have been proposed. Birnbaum suggested that (i) the values of Ql, ••• ,Qs may not be known to the statistician, (ii) the distributions of the Q., which are not necessarily the same for 1 all i, may not be known to the statistician, (iii) although the values and distributions of the statistics may be known, there may be no available overall test based on the values of the statistics themselves, or (iv) available methods may be difficult to apply. Usually, however, the statistician does have the relevant information concerning the statistics to be combined. If the available methods which are based on the values of the statistics themselves are difficult to apply, the statistician may prefer a simpler method based on the probability integral transforms if the loss in power is small. The arguments that combination procedures which do utilize the values of the statistics do not exist may be losing its validity as research in this area proceeds. Van Zwet and Oosterhoff [1967] and Oosterhoff [1969] proposed procedures for combining Ql and Q which are based on the statistics 2 themselves. These authors considered the cases when the distributions are either normal. or Student's-t distributions, but noted that their techniques could be applied to derive results if the statistics have distributions belonging to the one-parameter exponential family of 8 distributions. They proposed three overall tests and compared them with Fisher's test, finding none of the four procedures to be uniformly superior for the problems considered. Van Elteren [1958] considered linear combinations of two-sample Wilcoxon tests, giving attention to the properties of tests based on two different sets of weights. Bhattacharya [1961] compared three procedures for combining independent chi-squared statistics for sampling experiments. He found Tippett's method to be, in general, inferior to Fisher's method and to the overall test based on the addition of the s independent chi-squared statistics, but no uniform preference between the latter two methods was observed. The additi~n of independent chi-squared statistics for contingency table tests to obtain a test over several control groups is a well known procedure (see, for example, Hill [1971]). Mantel [1963] has proposed extensions of the Mantel-Haenszel procedure to various related contingency table problems. Schaafsma [1968] studied certain aspects of linear combinations of normally distributed statistics. Oosterhoff [1969] reported that Davies also considered procedures to combine normally distributed statistics, investigating B-optimal size-a tests to test the variance of the normal distribution. In the papers mentioned above, it has generally been assumed that the statistics to be combined have continuous distributions. In this dissertation consideration will be limited to the continuous case. Aspects of the discontinuous case have been discussed by Wallis [1942], Lancaster [1949], E. S. Pearson [1950], and Kincaid [1962]. 9 1.3. Introduction to Type DandType F Critical Regions As mentioned earlier, for certain multivariate testing problems type D and type F critical regions provide two locally most powerful tests found by maximizing different measures of the curvature of the power surface at _0 A. In this section the preliminary notions concerning the type D and type F critical regions are presented. 1.3.1. Type D Critical Regions Uniformly most powerful tests do not exist for many testing problems, but often locally most powerful tests can be found. For A ~ Ao , where A is a one-dimensional testing H : o parameter, the locally best unbiased critical region has been described by Neyman and Pearson [1936]. This so-called type A region of an unbiased test is found by maximiZing the curvature of the power A = Ao , for a fixed size of the test. at the point A = -0 A Generalizing to a multiparametric testing problem, H: o versus HI: curve A - ~ A where -0 ~ .... is a p-dimensional vector, it is reasonable to attempt to find the locally best unbiased region by maximizing the curvature of the surface of the power function at the point A = A • - -0 Among the class of all unbiased tests of a given size a, the test which has the critical region found by simultaneously maximiZing the curvature at every cross-section at the point A = _0 A does not exist. _ Pearson [1938] proposed a type C region, for the case p = 2. Neyman and A type C region for a testing problem is defined to be the region with the best local power along a given family of concentric ellipses with the same shape and orientation. If the concentric ellipses are circles, then the type C region is said to be regular; otherwise it is nonregular. How- ever, to construct a type C region it is necessary to know the relative 10 importance locally of type II errors, and this is not usually known. Furthermore, type C regions are not in~ariant which are one-to-one and twice differentiable. under transformations Therefore, transforma- tions of regular regions may be nonregular. To overcome these objections, Isaacson [1951] proposed a type D region, which is defined as the unbiased critical region of size a which maximizes the Gaussian curvature of the power surface at the point A = A. -0 The type D region can be constructed without knowledge of the type II errors and is invariant under one-to-one, twice differentiable transformations. Geometrically, a type D critical region provides intuitively appealing justifications for its adoption as an extension of a type A critical region. First, the type A critical region maxi- mizes the curvature of the power curve at the point specified in the null hypothesis and the type D region maximizes an average measure of the curvature of the power surface at the point specified in the null hypothesis. Also, since a type A region maximizes the curvature of the power function in an infinitesimal neighborhood of A = A , the o horizontal chord intersecting the power curve at a height a + £ is minimized among all chords so constructed for unbiased tests of size a. Similarly, among all unbiased critical regions of size a, the type D critical region minimizes the area of the region on a horizontal plane at a height of a + £ which is defined by the intersection of the plane and the power surface. For simplicity, the definition of a type D critical region will be presented in the two-parameter case. number of parameters is immediate. The extension to any finite Adapting the notation of Isaacson, for any region w in the set of all possible critical regions for a test, 11 the power function is given by B(~lw) where E= = Pr(U E wl~) (U 'U ) is a sample point in a 2-dimensional sample space. 1 2 Here, each U is distributed as a beta with parameters b and a and noni centrality H : o e = f(A-A ) .. ..0 for some function f. A = ..A0 is equivalent to testing H: 0 Notice that testing = 0, A A ..0 in each instance is that equality does not hold. if the alternative If the criteria used in determining a type D region are invariant under translation, there = ..O. is no loss in generality in assuming that ..0 A Thus the size of the test associated with the critical region w is given by B(A..0 Iw) = B«O,O)' Iw) = Pr(U.. E wl(o,O)') • Now define Bi = Bi(w) = a[B«A l ·A 2), Iw)] a Ai i (A ,A ) 1 2 = assuming that the derivatives exist. 1,2. (0.0) a2 [B«Al,A Z)' Iw)] a Ai a Aj = i,j = 1,Z, (Al,A ) Z = (0,0) A necessary condition for unbiased- ness is that (1. 3 .1.1) The Gaussian curvature of the power surface at A = ° is given by = det(B..w) • the determinant of B , where ..w 12 B ...w = [B ll (w) B (w) 12 The restriction (1.3.1.1) guarantees that a relative extremum exists at A =~. In the two-dimensional case, the extremum is a minimum if the Gaussian curvature is positive and if the curvature in all directions is positive, i.e., if det(B ) > 0 and B.. > 0, i ...w J.J. ~ 1,2 (Taylor and ~funn [1972, p. 227]). Notice that these two restrictions are necessary and sufficient conditions that B be positive definite. ...w = ditions that B l is unbiased. B 2 = 0 and ~w Together, the con- is positive definite imply that the test The type D critical region maximizes the Gaussian curva- ture, det(B ), among all regions w which provide unbiased tests of a ...w specified size. Since the Gaussian curvature is invariant under trans- lation, the above translation of ...0 A to 0 is not restrictive. To sum- marize: Definition 1.3.1.1. (Isaacson (1951]). A region w o is said to be an unbiased critical region of type D for testing the null hypothesis that A = 0 if: (i) B(Olw ) ... 0 = a; (1. 3.1. 2) = 1,2; (1.3.1.3) (iii) B is positive definite; ...w (1.3.1.4) (ii) B. (w ) = 0, i 0 J. o (iv) det(B ...w o ) > det(B ) for any other region w - ...w (1.3.1.5) satisfying (i) - (iii). Returning to the discussion of the geometrical interpretation, consider the power surfaces of size-a unbiased tests of a hypothesis Ho 13 and the set of regions on a horizontal plane at a he.ight of Ct + E which are defined by the intersection of the horizontal plane and the power surface. It will be shown that' the region which is formed by the inter- section of the plane and the power surface of the test based on the type D critical region has minliual area among all regions in the set. Consider the Taylor series expansion of the power function around O. Ignore terms of third and higher orders and assume that the derivatives of third order exist. Then If Al and A are in an infinitesimal neighborhood of Z excluded flom the expansion are negligible. definite, E = ~o = ~A'B ellipse defined by 0 TIO/ (det(~w)' mized. A > O. - -w- ~, the terms Since B is positive -w For a fixed value of E, the area of the = ~'~w~ is given by ,fO/ h Il B22 -Bf2 = For fixed E, this area is minimized when,det (~w) is maxi- By (1.3.1.5), w is the critical region which maximizes det(B ) o -w and thus minimizes the area of the ellipse. To extend the definition to any arbitrary number, p, of p~rameters, define ~ and 0 as pXl vectors, increase the range of the index i in (1.3.1.3) to p, and recognize Band B as pXp matrices. -w -wo The only point which requires further explanation is the verification that B -w being positive definite assures that a minimum exist~ at the o extremum point A -0 O. A pXp matrix _w B only if, for every nonzero p-vector is positive definite if and o ~, ~'~w ~ o > O. Assume that A 0 14 is not a point minimum of the power surface. '\ *, A Then there exists a point * ••• ').p) * in an infinitesimal neighborhood of ().l' = * ° such that B(). Iw ) < B(Olw ). - 0 - (1. 3. 1. 6) 0 By a Taylor series expansion, ignoring terms of third and higher orders and assuming that the third order derivatives exist, B().*lw) - 0 = B(Olw ) + 0 0 + * p ~(E~~= 1 ).~B .. (w ) A.. + 2E ~ ~~ 0 ~ .. 1 ~,J= * .. (w)A..) * A..B ~ ~J J 0 i<j < B(Olw ) - 0 by (1. 3.1. 3) and by assump'tion (1. 3.1. 6). The inequality in the last two lines implies that A. * 'B A.* < 0, which is not the case since B is -w0 -w0 by assumption positive definite. Thus the assumption that BO * Iw ) - does not attain a minimum at the point desired conclusion. =~ ~ 0 is false, which is the Therefore, B being positive definite is suf-wo ficient to show that the unbiased test achieves a minimum at A. O. 1.3.2. Type F Critical Regions The type F test is based on the concept of best average power, defined by Wa1d [1943] as follows: Definition 1.3.2.1. H: o ). =0 A critical region w for a testing problem versus HI: o ~ ~ ~ is said to have the uniformly best average power with respect to a family of surfaces R (A.) and a weight function - 6(~), c - if, for any other critical region w of the same size, f R ().) c ... Pr(w IA.) 6(A.) dA > f Pr(wlA.) 6(A.) dA 0 R ().) c - 15 where f $(A) dA is the surface integral of $(A) over R (A). _ Rc (A) If the surface is a sphere of radius r, so that A'A function c - - ~(A) - = r 2 , and the is a constant for all A, then it can be shown (in - Weyl [1939] and more simply in Roy ~ al. [1971, p. 25]) that This implies that if a deviation from the null hypothesis is equally likely in any direction, given ~'~ = r 2 , so that there is a uniform distribution over the points on the p-dimensional sphere of radius r with center at the origin, then the expected value of A'B A, the power - -wof the test at _A, is proportional to the trace of -w B. Since the trace of a positive definite matrix is positive, to maximize the expected or average power of the test subject to the restriction A'A only to maximize tr(B ) with respect to _w w~ = r 2 , one needs Following the notation of Neyman and Pearson and Isaacson, a critical region which provides the test with the best average local power will be termed a type F critical region. (Isaacson [1951] extended the notion of a type D critical region to include nuisance parameters and called such a region a type E To summarize, the following definition is offered: region. ) Definition 1.3.2.2. A region w o region of type F for testing H : o is said to be an unbiased critical A =~ against HI: A:fOifw o satisfies (i) - (iii) of Definition 1.3.1.1 and if ) > tr(B ) for every other critical region w -w o satisfying (i) - (iii). (iv) tr(B _w 16 Although both the type D and the type F tests provide tests which are locally most powerful on the average, by some means of averaging, to be consistent with Wald's terminology, a type D test will be referred to as a locally most powerful test while a type F test will be termed a test which is locally most powerful on the average. I . CHAPTER II TYPED CRITICAL REGIONS FOR COMBINING BETA-DISTRIBUTED STATISTICS 2.1. Introduction In Chapter I a critical region w was defined as an unbiased o critical region of type D for testing the null hypothesis that A = 0 if w satisfied four conditions. o In this chapter the concept of permissi- bility of a test is introduced and the form of the critical region of permissible tests is considered. By use of the concept of permissi- bility, it is next shown that a type D critical region does exist for the test procedure combining s independent statistics, U , ••• ,U ' where 1 s U has a beta distribution. i The exact form of the type D critical region is found under the assumption that the design is balanced in the sense that the noncentra1ity parameters are proportional with known proportionality constants. Under this assumption, expressions for find- ing critical values of size-a type D tests for combining two statistics are given. Finally, applications of these results to some F-distributed statistics which arise in analysis are mentioned. 2.2. Permissible Tests and Critical Regions There are two conditions which any reasonable test which combines statistics must satisfy. The first is that the combining test should depend on all of the statistics to be combined. The second condition, given by Birnbaum [1954, p. 562], requires that for every 18 point in the critical region of the test, every other point which represents a more extreme deviation be in the critical region. fr~m the null hypothesis must also The term "permissible test" is introduced to mean any test which satisfies these two conditions and is formally defined as follows: Definition 2.2.1. Let U., 1 1. ~ i ~ s, be s independent statistics for testing a null hypothesis H against an alternative HI. o Let W be the sample space, a subset of s-dimensional Euclidean space. If H is o rejected for large values of U., then a test of size a, 0 < a < 1, for 1. combining Ul, ••• ,U ' with its corresponding critical region w, is pers missible if: (i) w depends on each D., 1 < i _< s, in the sense that 1. - Pr(U = (Ul' ••• ,U., ••• ,U ) - 1 S E wlu.; U 1_ E W) is not constant for all values of U.; and 1. (ii) for any arbitrary point (ul' ••• 'u ) in w, every point in s the set u! > 1 - U., 1 1 < i < s; U 'EW} (2.2.1) is also in w. If H is rejected for small values of U., then the definition holds when o 1. the inequality in (2.2.1) is reversed. Since any reasonable combination procedure is permissible, permissibility cannot be used to distinguish effectively between competing procedures. Birnbaum noted, however, that more restrictive conditions which would narrow the number of choices cannot be specified without further assumptions. Having assumed that the densities of the 19 statistics to be combined all belong to the one-parameter exponential family, he considered properties of the acceptance regions of admissible tests, where admissibility is defined as follows: Definition 2.2.2. (Birnbaum [1954]). A test of a hypothesis is admissible i f there is no test of that hypothesis with the same level of significance which, without being less sensitive to any possible alternative hypothesis, is more sensitive to at least one alternative. Birnbaum [1955] showed that if the densities of the statistics belong to the exponential family, a necessary condition for the admissibility of the test is that the acceptance region be convex. Since the non- central beta distribution is not a member of the one-parameter exponential family with respect to the noncentrality parameter, Birnbaum's conclusion on the convexity of the acceptance region does not apply. However, the following characterization of permissible critical regions does apply and will be useful in subsequent derivations: Theorem 2.2.1. Let Ul, ••• ,U s be s independent statistics such that each U is distributed continuously over a finite or an infinite interval I • j j Let w be a size a, 0 < a < 1, critical region for a test ~(U) and assume that this test is permissi~le. ~(Ul' •• "Us) Also assume that the boundary between the rejection and acceptance regions is included in the rejection region. Then for any i, 1 ~ i ~ s, the boundary between the acceptance and rejection regions can be written 1 ~ j ~ s, where f i is a function which, in each of its (s-l) arguments, is non- increasing and is strictly decreasing at at least one point. 20 Proof. = If a 0 or a = 1, the f are not necessarily as specified; i l~ hence the restriction that 0 <a < Also, since the U.1. are continuous, . there is no loss in generality in assuming that the boundary is included in the rejection region. Since the test is permissible, it is clear from Definition ~(l) = 2.2.1 part (ii) that if, for any j, u(l) E > u U' j Wand u(2) = (ul""'u~, ••. ,u s J ) E (ul, •.• ,uj, •.. ,u ) s 4 w but w, it must be the case that r First let i = 1. Now assume that the boundary between the acceptance and rejection regions is increasing with respect to some argument, say u • 2 ~(3) E W, and u(4) = (ul,uZ, ... ,u ) s w where E U z < u z. In this instance the test is not permissible; therefore, the boundary cannot increase By the same argument, the boundary does not increase in u., J Next assume that the boundary does not depend on u. for 3 ~ j ~ s. some j, 2 J ~ j ~ s. Then part (i) of Definition Z.2.l is not satisfied. Therefore, the boundary must depend on each u ' 2 j boundary depends on each 2 ~ U., J 2 ~ j ~ Since the J j can be repeated for i j < s. s and is nonincreasing in each u., j ~ s, it must decrease in each u ' 2 ~ j This completes the proof if i ~ = 1. < s. Since the above argument = Z, ••• ,S, the theorem holds. 2.3. The Existence of a Type D Critical Region for Combining s Independent Statistics with Beta Distributions In this section the following theorem is proved: Theorem 2.3.1. Assume that U , 1 i ~ i ~ s, are s independent statistics and that each U has a beta distribution with parameters band a i i and 21 noncentrality parameter e • i eters are of the form e i Also assume that the noncentrality param- = A'ViA - - where each V., 1 -' -1 positive definite matrix and A is a pXI vector. ~ i ~ s, is a pxp If each U. is a 1 statistic for the testing problem defined by (1.1.1), then an unbiased critical region of type D does exist for the overall test of H which o Proof. To verify the existence of a type D critical region for the particular problem specified in the theorem, it must be shown that the four conditions (1.3.1.2) - (1.3.1.5) hold for this problem. condition specifies the size of the test. o -< The first Let u.1 < 1, 1 < i ~ s} - (2.3.1) and let w be a critical region, w c W, such that B(9!w) = a. Since the densities of the U. are continuous in W, this condition can be satisfied 1 for any predetermined value of a, 0 < a < 1. Once the form of the critical region is known, its exact boundary can be found for every such a. The first condition can thus be satisfied. To verify (1.3.1.3) for the particular problem of this theorem it must be shown that s eh Bi(w) = [a f IT (Su duh)/aA i ] w h=l h A=O I -- s eh = {Ljs=I[a f IT (Su duh)/ae. x cae./aA')]}1 w h=l h J J 1 where eh Su theorem. h 0, 1 ~ i ~ p, A=O (2.3.2) is defined in (1.1.2) with e h of the form specified in the For each i = I, ••• ,p and j = I, ••• ,s, 22 oOj/OA t act o(A'V.A)/oA = i - -J- i m'=l m=l A A , V(j»/dA m m mm' i A vq) = 2 [p m=l m m~ , (2.3.3) where v(j) = v(j) is the mm'-th element of V.• m'm nun' ~J For any i and j, (2.3.3) is zero if A = O. B. Therefore, (2.3.2) holds and ~ (w) = 0 for all i = l, ••• ,p for all critical regions w. The third condition to be satisfied requires that the matrix « [a2B (A- Iw) / (d A. dA.1 ,) ] I B = -w )) ~ A=O be positive definite. In the problem being considered here, the ii'-th element of B is -w 2 s 8h [d {! II (B duh)}/(dA.dA")]1 1 1 w h=l U h 8h Bu By expanding the densities (2.3.4) • A=O . in their series representations given h by (1.1.2), it can be shown (see Appendix A, Lemma A.I) that (2.3.4) equals r 02 ~'aA OA. L i 00 -8 /2 00 [E ... E l' r 1 e 1 =0 r =0 r (8 /2) 1/(r ! BCb+rl,a » l l 1 s -8 /2 x ••• xe x! s II s r (8 /2)' s s b+r -1 (u. j J w j=l a -1 (l-u.) j J } du.)] J I A=O 2 a = aA i dA i' /(r ! B(b+r ,a » s s s g[w;b;(8.,r.,a.), 1 1 ~ 1 ~ ~ s]l, J A=O (2.3.5) i (2.3.6) 23 = g[wjbj(8 i ,r.,a.), 1 1 where g in (2.3.5). 1 < i < s] is the term in square brackets -- C2.~.6) Upon differentiating, [ t ~ J ~=~ + °rs~x~ d) d8. ~ j=l d8 j dA i ' becomes <;,s. ~ " 1 '"'8 J= 0 j [ ~ X ---L-)] "" '" [d8. "" _0_ • (2.3.7) ol\i Ol\i' ;\=0 From (2.3.3), the first term of (2.3.7) is zero. Also from (2.3.3), (2.3.8) It remains to evaluate [dg/d8 j ] I '1~ j < s. Ignoring for the A=O moment the integral, consider r b+r.-1 a -1 e- 8j /2(8./2) j u. J (l-u.) j 00 _d_ J r J J r .! S(b+r . , a .) a8 j r.=O J J ~ J -8./2 e 00 = r J (-1/2) rj=O r r. b+r.-1 (8./2) J u. J a-1 (l-u.) j ..::..J~~~J_~_---=.J __ r.-1 -8./2 e j -A=O- r j ! S(b+rj,a j ) 00 + r J =1 J (r./2)(8./2) J J J b+r -1 a -1 u. j (l-u.) j J J A=O a.+b = [-1/2 + 1/2(---,\-) u j ] S~. (2.3.9) J since [(a.+b)/b]/SCb,a ) j J = 1/S(b+1,a.). J In light of (2.3.5) - (2.3.9), (2.3.4) can be rewritten (2.3.10) 24 s (j) Ej=l vii' where ¢. = -1 + (a.+b)u./b. J J J s 0 h~l ~SUh du h ) s IT (So du ) ¢j h h=l uh w v. J ~J v. ~ ¢j Therefore s B = E. 1 -w J= = s ~ x C (w) (2.3.11) j j=l -J s where C.(w) = J ¢. IT (So du ). h J w J h=l uh In order to show that _w B is positive definite, it suffices to show that the coefficients CJ.(w) > 0, 1 are by assumption positive definite. tion to the two-sample case. ~ s, since the matrices V. -J Let j = 1 and restrict considera- Then it can be shown that Cl(w) = J SO w ~ j u1 SO u2 ¢1 dU dU 1 2 whenever w is a permissible critical region. > 0 By rewriting (2.3.11), it can be seen that = (S O u l Su *) 0 l dUll dU 2 (2.3.12) where S0* denotes the density of a random variable U with a central l u 1 beta distribution with parameters (b+1) and aI' where [u2l denotes the lower bound of u ' and where [u11u2] denotes the lower bound of u 1 2 given u • Z If w is permissible, then by Theorem 2.2.1, the region of integration can be expressed in this manner as long as 0 < a < 1. Assume that [u11u2] = O. Then the inner integral of (2.3.12) becomes the integral over the range of positive density of the difference of two 25 density functions, which is zero. Since a l + b > b > 0, -1 + (al+b)ul/b is a strictly increasing function of .u . which attains its minimum at l u l = Oif o ~ ul It can, therefore, be concluded that ~ l- 1 I(u ) 2 = f [ullu z ] 1 SO u cl>l dU l ~ f 0 l 1 SO u with strict equality only if [ulluz] cl>l du 1 l = O. = f (Sou - Su0* )du l = 0 0 l l However, if 0 < ~ < 1 and w is a permissible critical region, [ulluz] must exceed zero for values U in some interval. z It thus follows that for any permissible critical region w of size a, 0 < a < 1, I(u ) > O. Z Similarly CZ(w) > O. Since SO U z is positive in Therefore, B is positive definite when s = 2. -w The demonstration that these coefficients are positive in the general s-sample case follows in identical fashion. Finally, it must be shown that there exists a size-a critical region w such that o s ~ det[~1' __ 1 for all other size-a critical regions w. form s ~i=l ~i x V. x C.(w)] = det(B ) _1 1 -w Since all matrices _w B of the Ci(w) are positive definite, det(~w) is a real positive- values function of B • -w Also, B is not a constant for all w since -w Ci(w) is not constant. It follows that det(B_w ) can be maximized and, therefore, the final condition can be satisfied. proof of Theorem Z.3.l. This completes the 26 Although it has been shown that a type D critical region does exist, it is not in general possible to find this region. Since the determinant function is not a finear operator, there is no simple method of finding (2.3.13) In Section 2.4 it will be shown that the type D critical region can be found if the design is balanced in the sense that V. _1 where the y. are scalars. 1 = y.V, 1 1- < i ~ s, The case of an unbalanced design will be considered in the third chapter. 2.4. The Type D Critical Region: Balanced Design In this section the following theorem, specifying the form of the type D critical region in the case OC a balanced design, is proved: Theorem 2.4.1. Vi - = Y.V, 1- If the conditions of Theorem 2.3.1 hold and if 1 < i < s, where the y. are known positive constants, then - - 1 the type D critical region of size a for the procedure combining wo = {(u1 , ••• ,u): s E1~-_1 Y.(a.+b)u./b > k'; 0 < u. < 1, 1 _< i _< s}, 1 1 1 1(2.4.1) where k' depends on a. Proof. Then by (2.3.11), if max[det(B )] -w w = det(V) P x [max(E~1=1 y.1 C.(w»J 1 W = det(V) Yi x [max(C )JP. w w = YiY, 27 Since Y and Ci(w) are all positive, C is positive. i w critical region is found by maximizing C. w Thus the type D This maximum can be found by applying the Neyman-Pearson Fundamental Lemma, stated here without proof. Lemma 2.4.1. (The Neyman-Pearson Fundamental Lanma). Suppose m + 1 given integrable functions g o ,gl, ••• ,gm are defined in an s-dimensional space. Consider the set of all regions w for which the following hold: 1 gi W s IT . 1 J= dU j = c., i = l, ••• ,m (2.4.2) 1 where the c. are m given constants. If w is a region which satisfies 1 0 (2.4.2) and if m ~ E:=l k.g. 1 1 g0 < L.1=1 k.g. 1 1 { go in w ' o (2.4.3) outside w ' o for m chosen constants k , then w has the property that 0 i s s 1 go IT duo 2-1 go IT duo J J j=l w j=l w 0 for any region w which satisfies (2.4.2). (Isaacson [1951]). s 0 s s 0 Let m = 1, go = (IT S )(E'- l y. ¢.), and gl = IT S ,so that c l = a. j=l U j 11 1 j=l u j If k l = (-L~=l Yi ) + k', s L -_ y.(a.+b)u./b > k'. 1 i l 1 1 then go 2- klg l if and only if Since the reverse inequalities also hold simultaneously, (2.4.3) holds for k l as defined above. Therefore, the type D critical region w , the critical region which maximizes C , is o w as specified in the theorem. This completes the proof. 28 2.5. The Type D Critical Region for a Size-a Combination Procedure in the Two-Sample Case In the previous section the form of the type D critical region w was found. o Given a, the problem is to determine k' in (2.4.1) such that f w o s II j=l o CSu. J du.) J = a. In this section this problem is handled in the two-sample case. For simplicity replace k' by k = k'b/[Y2(a +b)] so that 2 where f = [y (a +b)]/[Y2(a +b)]. l Z l The purpose is to find k such that for a given a, By Taylor series expansions of the terms in the integrand, several different expressions for the evaluated integral are found. All of the expressions involve double summations which have finite or infinite numbers of terms depending on which, if any, of the parameters aI' a , 2 and b are integers. It is of practical interest to use the expression which has the fewest infinite summations for a given set of parameters, and it can be seen that each of the different expressions found is the most useful in certain instances. At the onset it is useful to notice that the type D critical region is one of three different types, illustrated in Figure 2.5.1. The boundary between the acceptance and rejection regions is a segment of the line U l = (k-U 2)/f or, equivalently, U 2 =k - fUl' 29 For simplicity, in the following discussion let B. 1 let W be as defined in (2.3.1) with s = B(b,a.) and 1 = 2. Also, if x is a real posi- tive number, denote if x is an integer; otherwise. U U It 2 It ![\ \u1-(k-l) If I " -_,u1-(k-l)/f I \' ::w·· •• 0 , 0 0 u -k-f 2 " ..... u k' I II 0 0 k I '~ k , u l 1 •••• w •• 0 • • -.:: u -k-f 2 '. u o o I ~ f 'flL :: .w":: • •• 0 .. k l ....... ...... ..... o u 1 1 Csse 3 Case 2 Case I .. f 0 Fig. 2.5.1. Type D Critical Regions for Combining Two BetaDistributed Statistics Again for simplicity, some of the lemmas in Appendix A are used repeatedly without specific reference. In particular, Lemma A.2 justi- fies the interchanging of the order of iterated integration and Lemma A.3 justifies interchanging the order of summation and integration for this problem. The expansions and results of other lemmas are also used without reference. Expressions relating k and a are now found for each of the three types of area of integration. From Figure 2.5.1 it can be seen that in Case 1, 0 o~ (k-l)/f < 1, 1 ~ kIf, and 1 < k. ~ k-f < 1, Under the assumption that w is o such that these conditions hold, an expression for (2.5.1) which is 30 useful when b is an integer is first found. It is later shown that this expression is also useful if none of the parameters, aI' a , and b, is 2 an integer. Expressions which are more practical when aI' a , or both 2 are integers are also found. First write 1 - 1 w-w ° /30 /30 du du u u 1 2 l 2 (k-l)/f =1 - [1 o 1 /30 ( 1 0 ul k-fu -J /30 (I 1 /30 du )du . 2 l (k-l)/f u l 0 u2 1 + 1 (2.5.1.1) In the first term in brackets, the inner integral is unity, leaving where I(k_l)/f(b,a ) is the incomplete beta function ratio. l available by Pearson [1968]. Vol. 2 [1970, pp. 48-51].) evaluated by setting u b-l 2 (Tables For other references see Johnson and Kotz, The remaining term in (2.5.1.1) can be = (1-(1-u 2 » b-1 and proceeding as follows: 31 1 = i3if" 1 2 (b-1) * (b-1) E i=O . 1 l=!li a 2+'1 (2.5.1.3) The integral remaining to evaluate in (2.5.1.3) can be considered as the difference of two integrals, the first of which is (2.5.1.4) Using Lemma A.6 the second integral can be evaluated as follows: =f = a 2+i (b-1) * b 1 ' k-1 a 2+i+a 1+j E ( : )(_l)J (1 - - ) 13(a +j,a +i+1) 2 1 j=O J f a +i (l-k+f) 2 (2.5.1.5) Since both (k-f) and (k-1)/f are between zero and one in Case 1, neither (l-k+f) nor (1 - k~l) increases as its exponent increases. Combine (2.5.1.4) and (2.5.1.5) and apply Lemma A.5 to show that (2.5.1.3) is equal to 32 (b-I)* b-I i=!li a +i 1: (.) +' (I-k+f) Z 1 aZ 1 i=O x (b-I) * 1: j=O (2.5.1.6) These series converge in light of Lemma A.7. Substitute (2.5.1.2) and (2.5.1.6) into (2.5.1.1) and cancel terms where possible. Then i t is seen that (2.5.1.1) can be written (b-l) * b-I (_I)i a 2+i 1: ( , ) +' (l-k+f) i=O 1 a2 1 x * (b-l) a +j ( bJ-.l) (-1) j (1 _ k-fl) 1 1: S(a +j,a +i+l). 1 2 j=O If b is not an integer but either a l or a Z (2.5.1.7) or both are integers, another expansion of (2.5.1) may have fewer terms and thus be more useful. Consider again the final term in (2.5.1.1). It will now be evaluated by expanding the (l_u,)a i terms rather than the 1 before: a -1 x (2 i . )(_1)1 u~-l 1 terms as i\ 33 * (a 2-l) a -1 L (2 ) 1 = f3 l f3 2 i=O i (aI-I) * L j=O 1 x ! (k- f u )b+i d u • ub+j-l l l l (k-l)/f (2.5.1.8) Let y = fUl/k then apply Lemma A.5 to find that the remaining integral in (2.5.1.8) equals kb+i {(k/f)b+jr8f/k(b+j,b+i+l) - 8 _ / (b+j,b+i+l)]}. l l k (2.5.1.9) where 8 (c,d) = 8(c,d) x I (c,d) is the incomplete beta function. x x From the proof of Lemma A.7 the term in square brackets is bounded and even if the summation over j is infinite, that series converges. since k ~ 1 in Case 1, a 2 However, must be integral to assure the convergence of the series summed over i. By combining (2.5.1.1), (2.5.1.2), (2.5.1.8), and (2.5.1.9) and applying Lemma A.9, a second expansion for the integral over the critical region is found: 1 Il_(k_l)/f(al,b) - 8 8 1 2 a 2-l a -1 i (2.) (-1) kb+i i~O 1 b+i ~ x rf3 / k (b+j,b+i+l) - B1 _ 1 / k (b+j,b+i+l)] • f If a 2 gration of U l is not an integer but a ~nd U 2 can be reversed. l (2.5.1.10) is an integer, the order of inteEssentially the same steps are 34 integra~ performed to find that, for . 1 Il_(k_f)(aZ,b) - B S l 2 aI' (2.5.1.1) can be written a -1 1 a 1-1 (_l)i b+i i:O ( i )b+i (kIf) (a -1) 2 j:O * aZ-l ( j j )(-1) x [kb+j{Bl/k(b+j,b+i+l) - Bl_f/k(b+j,b+i+l)}l. (2.5.1.11) By Lemma A.7, the term in square brackets is bounded and the j-indexed series converges. If neither a 1 nor a Z is an integer, other expressions for (2.5.1.lZ) can be found. when (b-1) * They are, however, no improvement over using (2.5.1.7) = co. Three expressions, (2.5.1.7), (2.5.1.10), and (2.5.1.11), have been found which, given aI' a 2 , b, and f, can be used to find the k such that (Z.5.l.l2) equals some predetermined value a whenever w is of the a form considered in Case 1. It is not, in general, possible to solve for k in terms of a, so iterative solutions must be reLied on. The results of this section are formalized in the following theorem: Theorem 2.5.1.1. For each i = 1,2, let SO u. be the density of a random 1. variable U which is distributed according to a central beta distribui tion with parameters b and a .• Let w be a critical region, given by 1. 0 i where f is a known constant. If, for a given a € = 1,2} (0,1), (2.5.1.12) equals a and i f k':' 1 and kIf':' 1, then alternative evaluations of (2.5.l~12) are given by (2.5.1.7), (2.5.1.10) and (2.5.1.11). If either 35 (i) b is an integer and at least one of (a ,a ) is not an integer, or l 2 (ii) neither aI' a Z or b is an integer, then (Z.5.l.7) is the most use- ful of the three expressions. a 2 If b is not an integer but either a l or is an integer, then (2.5.1.11) and (2.5.1.10) are more useful than (2.5.1.7); expressions (2.5.1.10) and (2.5.1.11) both apply if a are both integers. l and a 2 If all three parameters are integers, all of the three expressions are applicable and all have a finite number of terms. 2.5.Z. Case Z In Case 2 there are essentially two subcases, as illustrated in Figure 2.5.1. Only the first alternative will be considered here, since the second subcase can be transformed into the first by resubscripting and proceeding accordingly. "Case Z" will now be used to refer specifi- cally to the first subcase, where 0 ~ (k-l)/f < 1, 0 < kIf 2 1, and 1 < k. In this case ! SO SO w l 0 u U (k-l) If z duldu Z = 1 - [! 0 kIf + ! (k-l)/f SO ! u 1 SO U l 0 SO ! u l 0 k-fu 1 SO u 2 z dUZdu l duZdu l ]· This expression is exactly (Z.5.l.l) except that kIf replaces 1 as the upper limit of the outer integral in the last term. The first term in brackets was evaluated in the previous section, so it remains only to evaluate the second term in brackets and combine terms. Since the techniques of integration used do not differ from those in the previous section, the details are omitted. 36 Corresponding to (2.5.1.7) is the expression 1 (b-l) * b-l (_I)i a 2+i Il_k/f(al' b) + - S - r (i) -; (l-k+f) S1 2 i=O a 2+1 In Case 2, 0 ~ (2.5.2.1) x 8 (I-k+f) (a 2+i+l, a 1+j)· 1/ a +j 1 (k-l)/f < 1 so (1 -(k-l)/f) does not increase with j. However, kif ~ 1 so 1 ~ (l-k+f) and thus (l-k+f) may increase with i. By Lemma A.a the i-indexed series converges anyway. Corresponding to (2.5.1.10) and (2.5.1.11) respectively are 1 I 1 _(k_1)/f(a 1 ,b) - Sl S2 (a -1) 2 i:O * ( a 2-1 (_l)i i )b+i b+' k 1 (aI-I) *a -1 r (1 ) (-l)j(k/f)b+j ~1/k(b+i+1,b+j) x j=O (2.5.2.2) j and (aI-I) r *a i=O . -1 1 . (1. )(-:) (k/f)b+1 1 b+1 x (2.5.2.3) Lemmas A.a and A.7 can be used to show the convergence of these series when kif ~ 1 and k > 1. To summarize: Theorem 2.5.2.1. Let U , SO , and w be defined as in Theorem 2.5.1.1. i u 0 If, for a given ex. € i k ~ (0,1), (2.5.1.12) equals ex and i f 0 ~ kif < 1 and 1, then alternative evaluations for (2.5.1.12) are given by (2.5.2.1), (2.5.2.2) and (2.5.2.3). None of these expressions restricts any of the 37 parameters to integral values, although (2.5.2.1) has the fewest terms if b is an integer and at least one of a (2.5.1.2) has the fewest terms if only a has the fewest terms if only a 1 1 and a 2 2 is not an integer, is an integer and (2.5.1.3) is an integer. As before, iterative solutions of k must be used since k cannot be found explicitly for a given a. 2.5.3. Case 3 In Case 3, 0 ~ k ~ 1 and 0 ~ kif ~ 1. Also in this case, the integral in (2.5.1) can be written I-kif f o (k'-l)/f f o where k' = (l-k+f). (2.5.3.1) If (2.5.3.1) and the term in square brackets in (2.5.1.1) are compared it is seen that they are identical except that 38 (i) U.1 has a beta distribution with parameters . band a beta distribution with parameters a by k'. i i while Y.1 has a and band (ii) k has been replaced Let o 0 . 8 8 du du ::: I. [k;(b,a ),(b,a ),f] 2 u u 1 1 1 2 1 2 where Ii [k;(b,a ),(b,a ),f] denotes an expansion of the integral if 1 2 w ::: wei) o 0 is a critical region which is covered in Case i, given the parameters of the beta distributions and f. Then the above comparison of (2.5.3.1) and (2.5.1.1) reveals that The appropriate substitutions in (2.5.1.7) and (2.5.1.10) give rise to the folloWing expressions, which are valid if the critical region is of the form considered in Case 3: (a -1)* 1 1 - 8 B 1 2 2 ~ i:::O , a 2-1 (_1)1 b+i ( i )b+' k 1 * (aI-I) a -1 , , t (1 )(_l)J(k/f)b+J 8(b+j,b+i+1); x x b~l j=O x (2.5.3.2) j j=O (b-:-1) (-l)j(l _ k-1)a1+ J j f ~f/(1_k+f)(al+j.a2+i+l) (2.5.3.3) 39 Since 0 ~ k ~ 1 and 0 ~ ~ kIf 1, neither a 2 nor a l in (2.5.3.2) is restr,icted to integer values. ' However,. since (l-k+O and (1-(k-1)/f) may exceed one, b is restricted to integer values in (2.5.3.3). Although (2.5.1.11) also has a corresponding expression in Case 3, it also requires that b be integral. What is needed is an expression which will improve on (2.5.3.2) when aI' but not a , is an integer. 2 It is given by (aI-I) * 1 1 1 2 L - 8 8 i=O (a 2-l) x *a -1 (2 L j=O . )(_l)J kb+j S(b+j,b+i+l). (2.5.3.4) j The results of this section are now summarized: a Let U., S , and w be as defined in Theorem 2.5.1.1. Theorem 2.5.3.1. ~ If, for a given a o~ kIf ~ E u 0 i (0,1), (2.5.1.12) equals a and if 0 cable only if b is an integer. l 1 and 1, then alternative expressions for (2.5.1.12) are given by (2.5.3.2), (2.5.3.3), and (2.5.3.4). that a ~ k~ and/or a 2 Expression (2.5.3.3) is appli- Neither (2.5.3.2) nor (2.5.3.4) requires be integral, but, for nonintegral b, if a 2 but not a is an integer, then (2.5.3.2) has the fewest number of terms while if a l but not a 2 is an integer, (2.5.3.4) has the fewest terms. As before, k cannot in general be solved for explicitly in terms of a. 2.6. The s-Sample Case In Section 2.5, discussion is limited to the two-sample case. As the number of statistics to be combined increases, the number of cases to be considered increases and the mathematics of evaluating l 40 the integrals becomes increasingly cumbersome. For small values of s, say s = 3 and s = 4, it is probably not unreasonable to seek solutions, although none are sought here. As s increases much further, the labor involved in evaluating the necessary integrals becomes extremely tedious. However, since the overall statistic for combining s independ- ent statistics Ul"",Us is where f holds. i = yi(ai+b)/b, it can be shown that the Central Limit Theorem The approximate normality of Z will be studied in Chapter IV. 2.7. A Note on Applications As noted in Chapter I, statistics with F distributions can be transformed into beta-distributed random variables. Therefore, the results of this chapter can be applied to the statistics discussed in Section 1.1. If, for example, there are s independent Hote11ing's T statistics to be combined, where (N i - p) T:/[(N.-I)pJ,i ]. ]. 2 = 1, ••• ,s, has an F distribution with p and Ni-p degrees of freedom, then it follows from the material in Section 1.1 that U.]. = [T~/(N.-1)J/ ]. l. 2 [l+T /(N -1)] has a beta distribution with parameters p/2 and (N -p)/2. i i i The results above can be applied by setting b and calculating f accordingly. = p/2 and a i = (N -p)/2 i In many actual experiments the number of variates may be quite small in comparison to the sample size so that b = [p/2] < [(N -p)/2] i = ai • If this is the case and s = 2, then for integral b the expressions derived in Section 2.5.1 - 2.5.3 which have summations extending to (b-1) have fewer terms than the other expressions. In particular, if b = 1, then b - 1 = 0 and expressions 41 (2.5.1.7), (2.5.2.1), and (2.5.3.3) have only one term. instance, k can be solved for explicitly in (2.5.1.7). In this Of course, for other sets of parameters the other expressions presented in Sections 2.5.1 - 2.5.3 may have fewer terms and thus be preferable. The results of this chapter also apply to the other statistics mentioned in Section 1.1. Again, if s = 2, the most useful expression for finding k, given a, can be chosen on the basis of the parameters of those statistics. CHAPTER III FURTHER CONSIDERATIONS IN THE LOCALLY OPTIMAL COMBINATION OF BETA-DISTRIBUTED STATISTICS 3.1. Introduction In this chapter the theoretical results of Chapter II are extended. In Section 3.2 a test which is locally most powerful on the average is found in the case of an unbalanced design; the locally most powerful test based on a type D critical region is in general not avai1able for this case. In Section 3.3 it is shown that when there are a large number of small-sample statistics to be combined, the locally most powerful test found in Chapter II is more efficient than Fisher's test according to Bahadur asymptotic relative efficiency, for local a1ternatives to the null hypothesis. It is also shown that, for local a1terna- tives, the type F test is more efficient than Fisher's test according to an average measure of Bahadur asymptotic relative efficiency. This result is of particular interest in light of the work of Littell and Folks [1971,1973] which was discussed in Chapter I. In the final section several other extensions of the results of the second chapter are considered. 3.2. The Type F Critical Region: Unbalanced Design The case of an unbalanced design is considered in this section. It is assumed that,under H , the s independent statistics uI, .•. ,U to s 1 43 be combined have noncentra1 beta distributions with noncentra1ity It is no longer assumed, however, that the design matrices ~i are proportional. It is known from Theorem 2.3.1 that a type D critical region does exist, but it is not in general possible to find the type D critical region since it is not in general possible to find where Band Ci(w) are defined in (2.3.11). -w Although it is not always possible to find the type D critical region in the case of an unbalanced design, it is possible to find the type F critical region, which provides a test which is locally most powerful on the average, as discussed in Chapter I. From Definition 1.3.2.1 and (2.3.11) it follows that the type F critical region is given by s max[tr(B)] = max[tr(L =l ~i x Ci(w»] i -w w w = s max[L '=1 tr(V.) x C.(w)] W where Y i = tr(V.). -1 -1 1 1 This maximum has already been found in Section 2.4 and the desired critical region appropriately redefined. W o is defined by (2.4.1) with Y i The results of Sections 2.5 - 2.7 also apply. If the design is indeed balanced, so that V. -1 then = o.1 V, 1 - ~ i ~ s, 44 s max[tr(B )] = max[E _ tr(V ) x C.(w)] i- l -w -i 1 w w = tr(V) x max[E~=l 0i Ci(w)]. w s Since Ei=l 0i Ci(w) is exactly the term maximized in Section 2.4 to find the type D region, the type F region is identical to the type D region if the design is balanced. It is of interest to attempt to find other combination procedures which are related to the type D and type F tests. In order to investigate this issue and to further compare the type D and type F tests, it is useful to consider the geometrical interpretations of the characteristic roots, ~1(w), ••• ,~ p (w), of B. -w Restrict consideration to the two dimensional case (p=2) and let ~1 and ~2 be the two (p X1) characteristic vectors associated with roots ~1 (w) and ~2(w). Assume that ~l(w) > ~2(w) and that ~1 and ~2 are normalized so that V'V -i-i = 1, i = 1,2. The vectors form a basis of a two-dimensional Since B _w is positive definite symmetric, that basis Euclidean space. is orthogonal (Graybill [1969, p. 47]). nates in terms of the usual axes. Let A be a vector with coordi- Now consider the ellipse, mentioned in Section 1.3.1, A'BA=E, ,. ,. _w- (3.2.1) defined by the approximate intersection of the power surface and a cross-sectional horizontal plane at a height of a + E. First let A be a vector along the ':l-axis, so that A = q1 ~l' where q1 is defined so that (3.2.1) holds. Then 45 since ~w ~l q2 = ~l(w) = '£7~2(w). ~l' whence ql = '£7~1(w). Similarly, if ~ = q2 ~2' By definition'~l(w)·~ ~2(w), and therefore ql ~ q2' implying that the larger the characteristic root, the shorter the axis of the ellipse in the direction of the associated characteristic vector. This relationship is pictorially displayed by superimposing the ellipse (3.2.1) on the ~1~2-axes, as shown in Figure 3.2.1. V ~2 Fig. 3.2.1. The Ellipse Formed by the Approximate Intersection of the Power Surface and a Horizontal Plane at a Height of a + £. Notice that the type D critical region is given by max[det(B )] w -w = max( p IT w i=l ~.(w» 1 since the determinant of a matrix is equal to the product of its p characteristic roots. By maximizing p IT iel ~i(w), one minimizes IT '£7~i(w) for fixed £, and, therefore, the type D critical region i=l provides the test which minimizes the product of the lengths of the axes of the ellipse A'B A. --~ On the other hand, the type F region maximizes tr(B ~i(w). _w ) = L~-l 1- For fixed p, the type D critical region is found by maximizing the geometric mean of the characteristic roots, 46 and the type F critical region is found by maximizing the arithmetic mean of the characteristic roots. In the literature there is one other combination of the characteristic roots which is in common usage, one which would provide a minimax test by minimizing the longest axis of the ellipse A'B A. --~ This test would maximize the curvature of the power surface along the axis where the curvature is the least by maximizing the minimum root. However, for this test it is necessary to find max w (Roy ~ {- t~'(L~=l-~imin X'X A x Ci(w»~ J.l (3.2.2) al. [1970, p. 26]), which is not in general possible in the case of an unbalanced design. If the design is balanced, so that !i = Yi~' 1 ~ i ~ s, then this minimax test coincides with the type D test. This can be shown by rewriting (3.?2) as s max (Li=l Ci(w) y.) x min 1 A w -[ A'V A] _ - A'A -- • (3.2.3) The second term in (3.2.3) is independent of w, and the solution to the first term is exactly the type D critical region which was found in Section 2.4 for the balanced case. 3.3. Bahadur Efficiency In this section it is shown that, for local alternatives to the null hypothesis, if there are a large number of finite-sample betadistributed statistics to be combined, the type D critical region provides a test which is superior to Fisher's test according to Bahadur relative efficiency. It is also shown that the type D critical region 47 provides a test which is superior to Fisher's test according to a measure of average Bahadur asymptotic relative efficiency. Littell and Folks [1971,1973] have shown that Fisher's test is in general at least as efficient as every other test when there are a small number of largesample statistics to be combined. However, other tests have been shown to be superior to Fisher's test for certain instances in which the two tests cannot be distinguished on the basis of Bahadur efficiency. Define the testing problem as H: A o For s = 1,2, ••• , = -0 A versus HI: A ~ A • -0 let T be a real valued statistic which depends on the s first s observations, {q1, ••• ,qs},of an infinite sequence. large values of T lead to rejection of H. s a sequence {T } is a s Assume that Also assume that the standard sequence, which is defined as follows: Definition 3.3.1. (Bahadur [1960]). A sequence {T } is said to be a s standard seguence for testing Ho if and only if {T } satisfies the s following conditions: (i) there exists a continuous distribution function F such that lim Pr o (T s < z) = F(z) for every z; s~ (ii) there exists a constant a, 0 < a < -2 log [1 - F(z)] 00, such that = az 2 [1 + 0(1)] (iii) there exists a function b(A), 0 < b(A) < for each A - ~ 00 , as z~; such that A, _0 l 2 lim pr A(IT s /s / - b(~) s-+oo _ If {T } satisfies these conditions, then s I > z) = 0 for every z>O. 48 A= A , -0 A '/: A , -0 is said to be the approximate slope of {T}. s To compare the asymptotic efficiencies of two standard sequences {T (I)} and {T (2)} with slopes s s cI(~) and c2(~)' respectively, Bahadur [1960] suggested ~12 = cI(~)/c2(~) ~12(~) if as a measure of asymptotic relative efficiency. > 1, the first sequence is more successful than the second; ~12(~) < ~12(~) If 1, the second is more successful than the first; and if = 1, the two sequences are equally successful according to this measure of efficiency. To compare the asymptotic relative efficiency of the test based on the type D critical region to Fisher's test for the testing problem which has been discussed in this dissertation, the normalized statistics are used. If UI, ••• ,U are s beta-distributed statistics to be com- s bined, the locally most powerful test was found in Chapter II to be S based on the statistic E = l[y.(a.+b)U./b] where a., b, and Yi are 1 1 1 1 i defined in Section 2.3. Let X.1 = (yi(a.+b)U./b) - E0 (y.(a.+b)U./b) 11 111 (yi(a.+b)Ui/b) - Y so that E (X.) = O. 1 0 1 i S are independent, Var o (E.1= 1 X.) = E S 1 i= Since the U., and hence the X., 1 1 Var 0 (X.). 1 1 Therefore, the normal- ized overall test statistic for the locally most powerful test, based on the type D critical region, can be written T (1) s s } s 1/2 = {E.1= I[X i - E0 (X.)] I[E.1= 1 Var 0 (X.)] 1 1 c s s 1/2 [Ei=l Xi]/[Ei=1 Varo(Xi )] • 1 Now let Y = -2 log f i U i eOz dz, where eOz is defined in (1.1.3). (3.3.1) Then 49 s Ei=l Yi is Fisher's overall test statistic for this problem. Define (3.3.2) The asymptotic relative efficiency of {T (I)} to {T (2)} is the s s asymptotic relative efficiency of the locally most powerful test to Fisher's test, as long as the conditions of Definition 3.3.1 are satisfied. The first condition can be satisfied since both T (1) and s T (2) are asymptotically normal, so that F = s standard normal distribution function. ~ in (i), where s-.oo ~ 00 o _ (Fisz [1963, pp. 206-207]). s-.oo E~=l Yi~ X~s' is the The Central Limit Theorem holds with respect to Ts(l) since the X1.. are bounded and lim Var s 2 lim Ei=l YI ai/[b(ai+b+l)] ~ (E~1.= 1 X.) 1. Under H , o and the asymptotic normality of T (2) is well established s (Johnson and Kotz, Vol. 1, [1970, p. 176]). satisfied when F = ~ The second condition is with a = 1 in (ii) (Bahadur [1960]). It remains only to show that (iii) holds for T (1) and T (2). s s Consider first the sequence {T (I)}. s s E =l EA(Xi)/s in probability as s i ~ 00. The Markov Law of Large Numbers (Fisz [1963, pp. 216-217]) assures this since s 2 2 L =l EA(Xi)/s ~ 0 as s ~ i o. s Under HI' E.1.= 1 x./s ~ 1. EA(X~) = 0(1) Furthermore, since 8. 1. = so that y.e, it can be 1. 50 (3.3.3) Also, 2 a.b 2 a.+b 1 1 = Li=l Yi (-b-) ----:2=----(ai+b) (ai+b+l) s s 2 = L I y. a./[b(a.+b+l)]. i= 1 1 1 (3.3.4) In light of (3.3.3) and (3.3.4) it can be seen that, for local alternatives to the null hypothesis, so that terms of order 82 and higher order are ignored, = s Notice that 0 < b l (~) -1/2 s 8/2[Var o (L i =1 Xi)] 112 - (3.3.5) • for Xi as defined in this problem. Next consider the sequence {T (2)}. s Under HI' L~=l Y/s ~ s Li=l EA(Yi)/s, again by the Markov Law of Large Numbers since, for this - 2 problem, EA(Y ) = 0(1). i In light of (1.1.2), the density of a non- central beta-distributed random variable U with parameters band a and i i 51 using the exponential expansion. Therefore, 1 1 s = Li=l f (-2 log f o u i (3.3.6) Substitute (3.3.6) into (3.3.2) and take the expectation to find, for local alternatives, that b (A) ~ s-1/2[8/2 L~ 1 E (Y. X.)] x [Var (L~ 1 y.)]-1/2. 2 1.= 0 1. 1. 0 1.= 1. Eo (Y i Xi) > 0 for all i since E0 (Y i X.) E0(X.) - E0(Y.) 1. 1. 1. - ..., - since as Ui increases both Xi and Y increase. i - (3.3.7) = E0(Y. X.) 1.1. Therefore, - b2(~) and > O. By combining (3.3.5) and (3.3.7), it is seen that the asymptotic relative efficiency of the locally most powerful test to Fisher's test, for local alternatives, is 2 E (Y. X.)] , 01.1. (3.3.8) as long as the limits taken exist. (In the context of experimental designs, there are general models where the limits taken in (3.3.8), in light of (3.3.5) and (3.3.7), exist. See Sen [1969] for details.) Also 52 s s Eo(Ei=l Y.1 E.1= 1 X.1 ) s s s Eo(Ei=l Yi Xi + Ei=l Ej=l Yi Xj ) i:!j s = E. 1 Eo (Y.1 Xi) ES E (X.) _ + E~ 1= 1=1 j=l Eo (Y.) 1 ~ J i:fj since Y. and X. are independent if i :f j and E (X.) = 0 for all j. 1 ~ J J Combine (3.3.8) and (3.3.9) to see that (3.3.10) There is equality whenever Xi and Xi are linearly related. Since Xi is a linear function of U. and Y. is a function of U. which is never 1 1 1 linear, the inequality is strict. There~ore, (3.3.11) This result implies that if there are a large number of finite-sample statistics to be combined, the test based on the type D critical region is more successful than Fisher's test according to asymptotic relative efficiency, for local alternatives toR • o It should be noted that the asymptotic relative efficiency based on approximate slopes can be misleading. It has been noted (for example, by Bahadur [1967]) that tests which are identical may not have identical approximate slopes so that does not occur if ¢12(~) approximate slopes. ¢12(~) is not 1. This dilemma is defined as the ratio of exact rather than Assume that the exact distribution function of T , s . * F , is such that (ii) in Definition 3.3.1 can be replaced by (ii) : s 53 (ii) * (1) lim F (z) = F(z) and (2) there exists a function f s s-+= on (0,00) into (0,00) such that, for any given sequence {v s } of positive constants where V ° < z < 00, 2/s log[l - F (v)] s Then c * (A) * * c * (~) 2 such that lim {v /s} s s-+= -f(z) [1 s f[b(A)2] is the exact slope of {T }. - s * cl(~)/c2(~)' then <l>12(~) ever, s =1 + = z, 0(1.)]. * If <1>12 (~) = if the tests are indeed identical. How- is often difficult to find, as it is for the sequence {T (I)}, and hence the use of approximate measures here. s Bahadur [1967] suggests, though, that for local alternatives to the null hypothesis, <l>12(~) is likely to be a good approximation of * <l>12(~). Since the results above are based on the assumption that the alternatives are local and considering the inherent properties of the test based on the type D region, it is reasonable to assume that the conclusion of the superior efficiency of this test over Fisher's test, for local alternatives, is not an artifact. If the design is not balanced, the final steps in (3.3.3) and (3.3.6), and hence the measure of relative efficiency given above, do not apply. However, similar results can be found for comparisons of the type F test and Fisher's test. Let Y i = tr(~i). Then the approxi- mate slope of {T (I)}, for local alternatives, is s (3.3.12) This result is analogous to (3.3.5) and follows immediately from (3.3.3) 54 and (3.3.• 4); the 8 1 have not been factored out since the assumption that 0i = Yi 8 no longer holds. Similarly, in light of (3.3.7), the approxi- mate slope for {T (2)} can be Written as s (3.3.13) For the unbalanced case, the Bahadur asymptotic relative efficiency, for local alternatives, is bi(~)/b;(~) in (3.3.12) and (3.3.13) respectively. where bl(~) b2(~) are given Unfortunately, this ratio is difficult to evaluate for the unbalanced case. ratio of the squares of the and Consider instead the average approximate slopes, where the average is found by integrating over the sphere A'A = r 2. The defini- tion (3.3. J.4) is introduced as a measure of average Bahadur asymptotic relative efficiency, where f b. (A)dA. A'A=r 2 ~ - Notice that this averaging is the same as that given in Uefinition 1.3.2.1. f To evaluate bl(~) and b2(~)' first apply (1.3.2.7) to find A'A=r2 6 .dA = f A'V AdA = oCr) x tr(V".) = oCr) x y. i A'A=r2 - _i_ -~ ~ where oCr) is a function of r. (3.3.15) Apply this integration to the right-hand side of (3.3.12) and recall (3.3.4) to find that, for local alternatives, 55 = s -1/2 s (o(r)/2)[Varo(~i=1 Xi)] 1/2 (3.3.16) • Similarly, from (3.3.13), b2(~) = s-1/2(O(r)/2)~s1. __l E (Y. X.) - 0 1 1 x [Var 0 (~~ 1 y.)]-1/2 1= (3.3.17) 1 for local alternatives. As before, in the present context o< It follows from (3.3.14), (3.3.16)-(3.3.17) b.(A) < 1 _ 00, i = 1,2. and (3.3.8)-(3.3.11) that locally, Therefore, according to the above measure of average Bahadur asymptotic relative efficiency, the type F test is superior to Fisher's test for local alternatives. As before, Bahadur asymptotic relative efficiency is not necessarily an accurate comparison of two tests. Furthermore, the processes of averaging used above may introduce other inaccuracies or difficulties in interpretation. Nevertheless, in light of the method of derivation of the type F critical region, the above result is of interest. 3.4. Extensions to Other Problems Locally optimal combination procedures for tests of hypotheses specifying the value of a vector of parameters have been found in this and the previous chapter. It is possible to find analogous procedures for combining certain tests of a null hypothesis of the form H • o· = 0, - (3.3.4.1) 56 even though the optimality of such procedures is not necessarily extended from previous results. The likelihood ratio statistic, A, for tests of (3.3.4.1) is often written U , where q,r, and t are appropriate parameters. q,r,t Let u(i) , 1 < i < s, be s independent statistics for testing a q,r,t i common hypothesis; q and r need not be indexed. Gupta [1971] showed that in the noncentra1 linear case (that is, when the alternative has q unit rank), U(i) is distributed as IT X.. where X.. has a central q,r,t 1.J j=l 1.J i beta distribution 130 x «r+1-j) /2, t/2) , 2 ij ~ j ~ q, and (1 - Xi!) has a noncentra1 beta density, so that XiI has density t 00 e -8 /2 i =e where e.1. L h=O -8 i /2 e.1 il )r/2+h-1/[2 hhl e(t /2,r/2+h)] i 00 L (8./2)h 130 Xi! 1. (t./2,r/2+h)/hl, 1. 1. l (See Gupta for details of For the present discussion it is sufficient to note is dependent on o e (1_x is the noncentrality parameter. Under H , X' g /2-1 h=O calculating 8 • i that 8~ Xii * ~'V.~ - _1_ has density 130 X. 1.1 for some appropriate design matrix V*•• ) -1. (r/2,t./2),1 < i < s. 1. -- Therefore, if (.) q (r;t./2) denotes the noncentral density of U. = U 1. t = IT X.. , ui - 1. 1. q,r, j=l 1.J where :' = [r/2,(r+I)/2, ••• ,(r+l-q)/2], then g e (r;t /2)= e ui - i -8 i /2 00 8 h I i L (-2) -hI g h=O . 0 ui (r/2+h,(r+l)/2, ••• ,(r+l-q)/2;t./2), 1. o -< u. < l. 1.- Recall that the type D and type F test statistics are of the form (3.3.4.2) 57 where Y is a weight dependent on the design matrix V. and So (b,a.) i _1 u 1 i is the density of a beta-variate U i wi~h parameters band a • i Analogously define the following as a test statistic for combining independent statistics U for testing (3.3.4.1): i g Yi g o U (r/2+l,(r+1)/2, ••• ,(r+1-q)/2;t./2) 1 i {3.3.4.3) 0 (r/2,(r+l)/2, ••• ,(r+l-q)/2;t /2) u. i 1 * where Y is a weight dependent on the design matrix _1 V.• If the funci o 0 tions gi = g (r;t./2) are known, then the statistic proposed in ui - 1 (3.3.4.3) can be calculated. known. For certain cases, the functions g~ are 1 See, for example, Gupta [1971] for the density of U when q,r,t q = 2,3,4, or 5. Notice that if q = 1 or r =1 , ~_qXr reduces to a vector and the likelihood statistic is distributed as a single betadistributed variate which is noncentra1 under the alternative hypothesis. In this case the results derived earlier apply directly. Assume again that the null hypothesis specifies the value of a vector rather than a matrix. Results similar to those in the previous sections and chapter can be derived if it is assumed that the statistics to be combined, X , ••• ,X ' have chi-squared distributions which are 1 s noncentra1 under the alternative hypothesis. chi-squared distribution with 2b i If X. has a noncentra1 1 degrees of freedom and noncentra1ity parameter 6 , then the density of Xi is given by i -6 /2 -x/2 b i 00 bi+j-l j 2j [e i e /2] ~j=O xi 6i /[r(b i +j)2 j!], x > 0, where r(a) = (a-I)!. It is easily shown that the type D and type F statistics for combining ~, ••• ,Xs are of the form 58 (3.3.4.3) where Y is a weight dependent on the design matrix _1 V. as before and i x2 (2b i ) denotes the density of a central chi-squared variate with 2b i degrees of freedom. Rowever, the chi-squared-distributed statistics which arise in real problems are often statistics used in the testing of (3.3.4.1). In particular, Sugiura and Fujikoshi [1969] have shown that the nonnull distribution of -2 log A, where A is the likelihood ratio statistic mentioned above, can be approximated by a linear combination of non-2 central chi-squared-distributed variates, up to order N where N is the sample size. For large samples, the distribution is approximately a chi-squared-distribution which is noncentral under the alternative. In that case the statistic (3.3.4.3) can be used to provide a combined test, with Xi = -2 log Ai' 1 is more complicated. ~ i ~ s. For smaller samples the result The overall statistic is similar to (3.3.4.2), with the central densities there replaced by the appropriate linear combinations of central chi-squared densities. Similarly, Siotani [1971] has given an asymptotic expansion of the nonnull distribution 2 -2 of the Rotelling-Lawley T up to order N o • Again the expansion is a linear combination of noncentra1 chi-squared variates which can be approximated by a single noncentra1 chi-squared variate for large samples. The extensions here follow in the same fashion as those involving combinations of likelihood ratio statistics. Finally, Pillai and Young [1971] have considered approximations 2 to the density of T for the central case. They have proposed that the o 2 central density of T be approximated by a generalized type II beta o 59 variate having density O~u~l. If research extending this result shows that the noncentral density of 2 T could be approximated by a noncentral beta-type distribution, a o procedure for combining s such statistics could easily be determined along the lines of the above extensions. CHAPTER IV APPROXIMATE CRITICAL VALUES FOR TYPE D AND TYPE F TESTS In Chapters II and III the type D and type F critical regions for tests of hypotheses specifying the value of a vector were developed. Formulas which can be used to tabulate critical values for two-sample problems were found in Section 2.5, and some tables of critical values are presented in Chapter V. Although it is, perhaps, reasonable to seek exact solutions for three-sample and four-sample problems, as the number of samples to be combined increases, both the number of cases to be considered in performing the integration and the complexity of the integrated results increase. Furthermore, the computer costs of creating the tables of critical values increase rapidly as the calculations become longer. It is possible, however, to use approximate rather than exact critical values in tests combining more than two samples. The type D s and type F test statistics are both of the form Li=l fiU i , where U is i beta-distributed with parameters band a., and f ~ an appropriately defined constant. s are bounded and Li=l Var(fiU i ) ~ 00 applies (Fisz [1963, pp. 206-207]). Therefore, for all real z i = (a.+b)Y./b with y. ~ ~ Since the s statistics U , 1 i ~ 2i 2 with s, the Central Limit Theorem s, 61 Pr(Z < z) = Pr . - [ ~~=l [fiUi - EO(fiU i )] . , < s l/Z [Ei=lVar (fiU i )] o ] Z -..;> ~(z) s-+oo - (4.1) ' s l/Z where Z = ~i l[f.U. - E (f.U.)]/[E. lVar (f.U.)] and where ~ is = 1. 1. 2 1. 1. 1.= 2 1. 1. S the cumulative distribution function of a standard normal random variable. It would be useful to determine values of s which are sufficiently large to provide good approximations in (4.1) so that the standard normal tables can be used to provide approximate critical values. Computer simulations were used to find the minimum 5 necessary to pro- vide good normal approximations for normalized type D and type F test statistics for several different testing problems. Different values of b, ai' and Y are used to define different combination problems. i The algorithm for the simulation is as follows: 1. Define the testing problem by determining b, and the range 2. Set s. 3. Generate s beta-distributed variates. (The beta variates were found by transforming two central chi-squared variates, say Xl and XZ' with Zb and Za degrees of freedom respectively. When Xl and X are so distributed, z B = XJ/(X l + XZ) has a central beta distribution with parameters band a. The variates Xl and Xz were found by summing the squares of, respectively, 2b and Za numbers which were generated from a standard normal population. These normally distributed numbers were generated by the PL-l version of program VARGEN 62 as installed on the University of North Carolina Computation Center's IBM 360/75.) 4. Normalize the sum of the s beta-distributed variates in order to calculate Z in (4.1). 5. Repeat steps 3 and 4 until a total of 5000 values, say Z., J 1 ~ j ~ 5000, are generated. 6. Perform the Ko1mogorov-Smirnov goodness of fit test at the .05 level of significance to measure the maximum distance D between the empirical distribution function of the {Z.} ~ and the standard normal cumulative distribution function and to test whether or not the values {Z.} are consistent ~ with a standard normal population. (The PL-1 Scientific Subroutine Package NDTR as installed at the U.N.C. Computation Center was used to determine values of the standard normal cumulative distribution function.) If D ~ 1.36//5000 = 0.0193, the hypothesis that the {Zi} were likely to have been drawn from a standard normal population is not rejected. 7. If D > 0.0193, increase s and go to step 3; otherwise, go to step 1 to define a new problem or else stop. In Table 4.1 the maximum distance between the empirical distribution function of the normalized variates Zi' 1 ~ i ~ 5000, and the standard normal cumulative distribution function is given for a number of testing problems. Y 1 a ••• = Ys. It is assumed that a If Y = Y, 1 ~ i ~ i 1 = aM s, Zi is not a function of the Yi's and hence there is no loss in assuming that Y i = Y = 1, 1 ~ i ~ s. 63 TABLE 4.1 VALUES OF THE KOLMOGOROV-SMIRNOV STATISTIC FOR CERTAIN SIMULATION PROBLEMS, ASSUMING CERTAIN PARAMETERS ARE CONSTANT FOR ALL SAMPLES b y a s D 1 1 1 1 1 1 3 4 0.0262 0.0154 1 1 1 1 2 2 4 5 0.0299 0.0152 1 1 1 1 1 1 8 8 8 7 12 20 0.0348 0.0207 0.0182 2 1 2 3 0.0178 2 2 1 1 8 8 7 12 0.0278 0.0145 3 1 8 8 0.0158 3 3 1 1 15 15 12" 16 0.0217 0.0136 Notice that for fixed b, as a increases the s needed to attain values of D less than 0.0193 also increases. This is not surprising since the beta distribution is symmetric if a = b (uniform if a = b = 1, be11shaped if a =b > 1) but is skewed if a ~ b. The parameter a is more likely to be larger than b in practical applications, and so the table has been prepared in that fashion. 1 ~ i able. ~ However, if ai = a and Yi = Y, s, then the parameters a and b are algebraically interchange- For example, from the table it can be concluded that if b = 2 and a = a = 1, 1 ~ i ~ s, s = 5 is the minimum s which is large enough i to provide a satisfactory approximation to normality, according to the criteria previously established. It can also be concluded, considering 64 the relative skewness of the beta density functions involved, that since s must exceed 12 if b =3 and a = 15 before approximate normality is achieved, s must be quite large indeed if a = 15 and b < 3. Table 4.2 indicates results of simulations when it is not assumed that the Y and the a are necessarily constant. i i r = (Y1 '···'Ys ) and = ~ (a 1 , ••• ,a s ). Let Other simulations indicated, as expected, that the approximate normality was very. sensitive to the inclusion of large a. when b was small. 1 While this table is i11ustra- tive of some effects of having varying Yi and ai' the exact parameters of the Y and a would have to be known before a reasonable guess at i i the minimum s which is large enough to produce good approximations to normality could be made. TABLE 4.2 VALUES OF THE KOLMOGOROV-SMIRNOV STATISTIC FOR CERTAIN SIMULATION PROBLEMS WITHOUT RESTRICTION ON THE CONSTANCY OF PARAMETERS b ...Y a s 1 12 0.0137 D 2 (1.0,1.2,1.4,1.5, 1.6,1.8,2.0,2.2, 2.4,2.5,2.6,3.0) 2 (1.0,1.5,1.75,2.0, 2.25,2.5,2.75,3.25) (1,4,5,7 8,10,11,14) 8 0.0207 2 (1.0,1.2,1.4,1.5, 1.6,1.8,2.0,2.2, 2.4,2.6,2.6,2.8) (1,2,3,4, 5,6,7,8, 9,10,11,12) 12 0.0189 2 ..1 (1,2,5,7, 8,10,13,15) 8 0.0145 ·I CHAPTER V TABLES OF CRITICAL VALUES FOR TYPE D AND TYPE F TESTS AND A NUMERICAL ILLUSTRATION 5.1. Tables of Critical Values for Type D and Type F Tests Tables of critical values for type F and type D statistics have been prepared for certain parameter values. Assume that U , i i = 1,2, has a central beta distrihution with parameters band a. and let 1 f = (a l +b)y l /[(a 2+b)y 2 ] where Y has an appropriate definition (see i Sections 2.4 and 3.2). for a f =1 Table 5.1.1 gives values ofk which satisfy = .01, .05; b = 1,2; a.1 = 1 (1) 10 (2) 30, i = 1,2; and (.5) 2.5. Incomplete tables are also given for b = 1.5. By comparing, for example (2.5.3.2) and (2.5.3.3) for Case 3, it is clear that there are generally more terms in the series to evaluate if b is not an integer but a chosen. l and a 2 are integers, for the range of parameters Since there are more terms to evaluate, the tables for b are substantially more expensive to compute. = 1.5 By the time the Case 3 runs for b = 1.5 had been completed it was evident that sufficient computer funds to complete the table were not available. Solutions for values of k given aI' a , b, and f were found 2 iteratively for the specified value of a, except for the few instances in which direct. solutions exist. Iteration started with an initial TABLE 5.1.1 CRITICAL VALUES FOR TWO-SAMPLE TYPE D AND TYPE F TESTS b = 1.0 ex = 0.01 al f 1 2 3 'I , 6 7 8 ., a2 10 12 III 1 .. 18 20 22 26 28 50 1.37'1 1.316 1.271 1.~31 1.77~ 1.733 2.29~ 2.2"5 2.205 2.771 2.720 2.682 2 1;0 1.5 2.0 2.5 1.68' 1.505 1.369 2.093 1.894 1.758 2.~C.7 '.300 2.1f7 2.'2r, 2.717 2.590 1.271 1.666 2.oe2 2.511 1.200 1.602 2.025 2.460 1.1"7 1.556 1.966 2.426 3 1.0 1.5 2.0 2.5 1.553 1.894 2.248 2.611 1.369 1.695 2.044 2.407 1.139 1.476 1.840 2.217 1.070 1.417 1.788 2.169 1.019 .982 .'53 .932 .915 .890 .80;0 .1143 .A37 .A~2 .A28 .872 .1160 .8;>5 .822 1.376 1.346 1.'23 1.306 1.292 1.271 1.257 1.'46 1.2'A 1.'31 1.226 1.~22 1.210 1.21~ 1.21' 1.75' 1.726 1.706 1.690 1.678 1.659 1.646 1.~36 1.6'8 1.622 1.617 1.';'13 1.609 1."~6 1.6C4 ~.136 2.111 2.092 2.078 2.06t; 2.048 2.036 ;>.026 2.019 2.013 ;>.oor. 2.004 2.001 1.99P 1.~96 II t.O 1.5 2.0 2.5 1.451 1.740 2.0 4 4 2.357 t.271 1.1~9 1.044 .974 .92:'1 .1185 .~56 .834 .1'117 .791 .773 .760 .7'50 .7 4 3 .n7 .7'12 .721$ .724 .721 1.545 1.~15 1.328 1.269 1.226 1.197 1.174 1.1~7 1.142 1.121 1.106 1.096 1.0n7 1.C91 1.075 1.071 1.067 1.0~4 1.~t;1 1.6 4 3 1.7~0 1.6~2 1.569 1.553 1.~26 1.506 1.49C 1.477 1.'158 1.44~ 1.~34 1.~27 1.420 1.415 1.411 1.408 1.40"' 1.~02 2.156 2.039 1.967 1.91 9 1.86~ 1.860 1.842 1.827 1.815 1.797 1.784 1.775 1.7~7 1.761 1.756 1.752 1.749 1.74~ 1.744 5 1.0 t.5 2.0 1.374 1.200 1.070 .974 .903 .851 .811 .7P.l .7~8 .739 .712 .694 .61\0 .670 .662 .~56 .6~1 .1'046 .643 .640 1.623 1.431 1.302 1.214 1.154 1.111 1.079 1.055 1.0~7 1.022 1.000 .985 .974 .91'.5 .959 .953 .9 4 8 .94r:, .941 .93'1 4 1.~85 1.687 1.5~3 1.4A4 1.431 1.393 1.365 1.'144 1.3~A 1.315 1.~~5 1.282 1.271 1.~h3 1.257 1.~52 1.24~ 1.2 4 1.241 1.~~q 2.157 1.Q58 1.841 1.767 1.718 1.683 1.658 1.639 1.623 1.611 1.r,93 1.580 1.~70 1.~h3 1.557 1.552 1.548 1.~44 ~.'5u, 1.~39 :o.~ 2.' 2.~ 1.2'5 1.5f3 1.918 2.2 11 9 1.107 1.523 1.958 2."01 1.'35 1.206 1.182 1.1'" 1.119 1.~" 1.673 1.651 1.618 1.595 2.173 2.149 2.12~ 2.096 2.077 2.~5~ 2.6~9 2.611 '.'83 2.563 1.077 1.4" 1.938 '.382 1.099 1.0~" 1.072 1.062 1.~77 1.'~4 1.~54 1.5'" 2.061 2.00;0 2.~40 2.033 2.5"9 2.5~e 2.~30 2."4 211 1.e5., 1.689 1.'~3 1.451 2.327 2.1"" 2.005 1.9011 2.eco 2.609 2.4~6 '.369 3.27~ 3.07~ 2.~~8 2.8"0 1 1.0 1.5 1.05" 1.0"7 1.5~8 1.533 2.027 :0.023 2.519 2.514 1.0"2 1.037 1.~2~ 1.~24 2.019 2.01~ 2.511 2.~C~ 1.053 1.034 1.007 .9S8 .'67 .'<;6 .9'5~ .953 .948 .941.6 .94~ .'37 1.~~0 1.466 1.445 1.430 1.'119 1.411 1.405 1.399 1.395 1.391 1.3=P 1.~e6 1.9~2 1.~09 1.~90 1.077 1.1I~7 1.1\59 l.e~3 1.p4e 1.844 I.A40 1.e~7 1."3"' 2.~67 2.~55 2.337 2.325 2.~15 2.308 2.'02 2.297 2.293 2.290 2.~P7 2.~~4 6 1.0 1.5 2.0 2.5 1.147 1.344 1.761 1.56'1 2.00n 1.802 7 1.0 1.5 2.0 2.5 1.271 1.460 1."'63 I. Q 74 8 1.0 1.5 '.0 2.!i 1.23'1 1.077 .~~3 .856 .723 .678 .1'.43 .7Pl .6'6 .~94 .~62 .539 .'\23 .480; .'512 .~02 .49'1 .490 .4~1 .477 1.4~3 1.223 1.094 1.002 .936 .8011 .8,,1 .1\24 .M3 .786 .761 .744 .731 .1"2 .7:4 .709 .704 .700 .6':'6 .f,'l3 1.583 1.392 1.266 1.181 1.1~2 1.060 1.049 1.0;>6 1.000 .')93 .97;> .957 .945 .9~7 .,.30 .925 .9?1 .917 .914 .911 1.772 1.~77 1.~55 1.376 1.322 1.284 1.256 1.23~ 1.219 1.206 t.186 1.17;> 1.162 1.154 1.148 1.142 1.138 1.135 1.132 1.129 , 1.205 1.053 1.0 I."; 2.0 2.5 10 1.0 1.~ 2.0 2.5 1.31~ 1.~32 1.019 .923 .e.Sl .796 .755 .723 .691\ .679 .650 .630 .1'.16 .606 .5'17 .'591 .586 .'581 .577 .'574 1.215 1.126 1.063 1.019 .986 .9"'1 .941 .9?6 .903 .667 .676 ;1l~7 .~60 .~~l.6 .8~0 .1\46 .843 • B4,':' 1.441 1.360 1.304 1.265 1.237 1.215 1.1"6 1.185 1.165 1.150 1.140 1.B2 1.12(, 1.120 1.116 loll? l.l~9 1.107 l.f83 1.608 1.557 1.522 1.49(, 1.476 1.4;0 1.448 1.429 1.416 1.406 1.3"8 1.'92 1.~07 1.3113 1.380 1.377 1.374 1.107 .9A2 .885 .811 .755 .712 .678 .652 .613 .601 .580 .565 .554 .~45 .538 .533 .528 .~2'+ .521 1.276 1.147 1.057 .9 9 2 .946 .911 .111\"; ./1<,5 .049 .1125 .11011 .7% .7~7 .71\0 .774 .770 .71:a6 .762 .71,;' 1. 4 69 1.344 1.2"'0 1.2n3 1.163 1.133 1.111 1.0q3 1.079 1.059 1.044 1.0:'13 1.025 1.019 1.013 1.0n9 1.005 1.00' .999 1.678 1.557 1.4eo 1.4~9 1.391 1.364 1.:'144 1.3~8 1.315 1.~96 1.282 1.<72 1.264 1.'58 1.253 1.249 1.245 1.242 1.240 .9::'2 .834 1.0~3 .9~9 .758 .69!1 .652 .616 .5P7 .564 .8"11 .752 .641 .r03 .774 .734 1.519 1.33n 1.2C3 1.116 1.n55 1.01~ .,:eo .9;;5 .Cj~f, .nl 1.6rA 1.495 1.371 1.290 1.235 1.196 1.167 1.145 1.128 1.115 1.30;~ 1.1~0 .530 .7OR .lIq"! 1.09~ 1.1e2 1.03~ .915 .817 .7~9 .679 .613 .5"4 .5"4 .~40 .504 1.317 1.14(, 1.019 .925 .1\55 .764 .733 .7] 0 .691 .1lO~ .664 1.4"4 1.279 1.151 1.0f3 1.000 .95'; .922 .1197 .877 .861 .838 1.618 1.426 1.302 1.219 1.162 1.122 1.092 1.069 1.052 1.030 1.017 .506 .690 .4 Q 9 .477 .6(,7 .Bln .P72 .EI';4 1.060 1.070 1.052 .479 .645 .822 1.CO~ .~77 • ~ (,1 .L+-1IB .~32 .621 .1\10 .992 .'la3 .8~1 .467 .460 .654 .454 .649 .449 .645 .P~l .1'57 .?1.47 .q u 3 1.0~5 1.050 1.046 1.042 .~~O .~38 .<,1 4 .794 .'177 .430 .607 .789 .972 .423 .602 .79" .967 .641 .441 ."'3A .S4r, .~37 .4lot" 1.0~" 1.~36 .418 .5"18 .a:" .7~O .777 .'%1 .96~ .'59~ .41~ .";91 .774 .958 0\ 0\ TABLE 5.1.1 (Continued) b = 1.0 a. = 0.01 al f 6 7 8 , .791 .712 .650 1.1"5 1.007 .890 .117'1 .802 1.257 1.09" .970 .7"6 1.37~ 1.201 1.07'1 .982 .916 .868 1.~0~ 1.321 1.1'15 1.109 1.0_9 l.OOE. .601 .70'1 .832 .97'1 .56?- .5~0 .1105 .'50 1'1 1.0 1.5 <'.0 2.5 1.119 .'le!\ .e72 .773 .1140 1.213 1.0S A .9~7 1.317 1.1"5 1.019 .926 1."2Q 1.2"5 1.118 1.030 .6"'+ .765 .P57 .966 .630 .707 .80F. .921 .~co .662 .767 .ae7 16 1.0 1.5 2.0 2.5 1.t,9Q 18 1.0 1.'1 2.0 1.0e~ 1 12 1.0 1.5 2.0 2.5 ~.5 2 .967 1.leo 1.032 1.2n 1.H4 1.3~6 1.le9 3 ./IF'0 .913 .9110 1.06~ .850 .8'5 .951 ."56 1.013 1.233 1,073 1.317 1.144 1.0~9 1.1~5 It 5 ._03 .383 .53" .(,1\9 .814' .376 .5211 .6'15 .1142 .370 .0;2_ .(,111 .e38 .362 .~2n .~17 .677 .e35 .E.74 .6 r ,Q .392 .';'10 ."95 .AS2 .3E.E. .725 .81'l .'118 .5(,0 .712 .1\68 .539 .(,27 .7:1S .861 .506 .600 .715 •et41 .'+7<) .577 .696 .825 .'+37 .'\'.5 .670 .802 .'+08 .522 .651 .705 .~"6 .~70 .358 .~06 .4 0 4 ... !O5 • 329 .463 .621l .7"" ."20 .757 .3"1 .472 .1'.09 • 7'~ 7 .334 ... 67 ."38 .773 .3"C ."713 .(,14 .752 .1;05 .~nl .~q~ .7"3 .740 .737 .!J65 .52~ ... 61 .5"2 .6"3 .75" .'+1P. .507 .61" .730 .386 ."83 .594 .712 • :'\33 ."43 ."'61 .693 .322 .435 .555 .677 .314 • :'\07 .'i95 ,"A" .5"6 .663 .771 .3"6 .632 .719 .820 • '.29 .42" .302 ."19 .2H ."16 .~"'5 .!ltl .'535 .£00'3 .6f~ .'177 .4.11 .516 .370 .452 .550 .654 .3"6 ... 33 .329 .275 .5~~ .f-41 .6~1 .302 ... 01 .5011 .£17 .2~0 .~34 .313 ... 09 .'i14 .f-23 .3'=:'& .699 .403 •• 78 .571 .672 ."94 .605 .3oC:C .49C .602 .392 .45£ .536 .626 .358 .421\ .514 .607 .~33 .313 .257 ."08 .""7 • .,93 .!~3 -.674 .'138 .495 .569 .654 ."30 .1f79 .51f3 .6111 .303 .4311 .509 .'i1l9 .3 .. 11 .409 .484 .566 .322 ."'71 .,,97 .642 .700 .7f-9 .545 .593 .655 .729 .502 .">53 .1;19 .6'" .538 .~C?5 1.:>75 1.111 22 1.0 1.5 1.062 .953 .C?A7 1.117 1.HO 1. 031 1.2"6 1.C1l4 .837 .1l71 .912 .961 .737 .772 .B15 .85!' .6:"6 .739 .7"" .591 .630 .679 .739 1.0SIf .9"5 .97f\ 1.10" 1.1"0 1.016 1.220 1.062 .8:12 .8/'.3 .e q 9 .941 .732 .763 .1101 .e45 .651 .6113 .723 .772 .C? 4 4 .828 .8"6 .81'7 .925 .72B 2.5 1.01f7 1.092 • 971 1.1"3 1.004 1.1911 1.044 28 1.0 1.5 2.0 2.5 1.042 .940 .~6~ 1.083 1.12" .994 1,179 1.030 .8:>5 .822 .845 .871 .9nl ."37 .6"6 .783 .9:'11 ."65 .595 .7'1:> .89'1 .662 .704 .7:"9 .11<'2 2.5 1.037 .937 1.e74 .960 1.117 .'lIlF. 1.103 1.019 30 .50'1 .625 .767 .916 .512 .57:> .1;49 .7 .. 0 .9'9 .9 11 6 30 1.0 1.5 2.0 2.5 28 .554 .610 .603 .769 1.0'~' .e~o 26 .606 .659 .725 ./l07 1.?~4 .878 .912 Z.. .670 .719 .7112 .858 ~.o ~.O 22 .750 .7'H .655 .926 .e~2 26 1.0 1.5 20 .81'1 .905 .99~ 2.5 18 .(,16 .679 .760 .956 .743 .793 .e33 • 893 ~.o 16 .7~9 .8 .. 3 21& 1.0 1.5 1- .6eD ."5~ 2.5 12 .no 1.072 1.13'1 ~.o 10 .815 .pe5 .971 20 1.0 1.5 a2 .1'.1\7 .793 .5~1 .6;'3 .717 .1fF,7 .521 .592 .;'(j~ .57~ .~(,3 ."65 .~80 .700 .5·~ .703 .""2 .570 .6'0 .4~0 .298 .31\2 ... 76 .,,>75 .2116 .27" .~c)3 .302 .372 .4.,,, .542 • 314 .2'l3 .~71 .3"~ ." n!1 • ~D 3 .633 .697 .596 ."64 .41'.0 .506 .5.;7 .639 .5e6 .620 .663 .715 .533 .569 .615 .671 .1f90 .5211 .577 ."37 .'15 4 .'I'J1f .51f7 .610 .423 .'167 .5'23 .589 .376 .424 .48" .557 .3 .. 1 .394 .460 .536 .442 .'\20 • 581 .612 .G5r .69" • 528 .560 .601 .650 .4115 .519 .5/'.2 .;'15 .449 ." 05 .4H .'156 .50"> ,564 .370 .413 .467 .551 .33 .. .381 ..... 0 .508 .307 .358 ... 21 .492 .2P6 .340 .4r7 .269 .327 .789 .829 .6"6 .676 .711 .753 .4;;0 ... 71 • 72 If .750 .780 .1115 .643 .669 .701 .736 .577 .605 .639 .67" .52'1 .553 .589 .£63 .4111 .'\11 .550 .5'6 .441f ."77 .518 .567 .414 .366 ."03 ."52 ."09 .329 .371 .42" .4P5 .302 .347 .404 •• "'1\ ,2"1) .'62 .311f .377 ..... 6 .477 .505 .'i39 • .,81 .441 .'170 .5n6 .551 .7~6 .721 .745 .772 .E03 .';93 .640 .1i64 .f,'l2 .726 .574 .599 .630 .666 .seo .521 .547 .580 .619 .c;~q .5~1 .587 .~4e ... 91 .54" .410 .4 .. 1 .4110 .527 .%2 .3<)5 .439 .491 .325 .362 .410 .466 .n€> .373 .46' .568 .~116 ."67 .297 .~F.7 .3119 ."48 .4~9 .5r9 .3~1I .3n9 .""0; .275 .31 A .373 .4~4 .550 ."72 .293 .66e .(,12 .2% .389 •• 97 .605 .277 .366 .463 .565 .269 .360 ... 58 .5S9 .£';2 .264 .3103 .u:r:l .522 .3'" .0;02 .~-:2 .4«'1 .">~5 .~C.2 .255 .337 .1026 .518 .2"" .332 ."22 .2103 .51~ .".11 .237 .312 .231 .3ce .3'lC .1f7e, • ~~,c .~~4 .437 .527 .277 .3 4 2 .4111 .'iCO .264 .332 .410 .493 .253 .32" .41l7 .2 .... .317 .3911 .4113 .255 .316 • 3P.7 .463 .2 .... .305 .361 .4sa .23"> .301 .375 .453 .227 .295 .371 .2"" .303 .237 .29" .361 ." 32 .227 .267 ,355 .427 .21'1 .231 .283 .221 .275 .338 ... 05 .257 • :'\04 .3(,1 .t42'" .~61\ ,431:1 .2"3 .;"l?2 .351 .4:6 •.~"14 ."10 .4~O ,lot ~ .. .~60 .?96 .3.,\ .3'5~ .... 5 .404 .'!33 .~q4 .,+7Q .'ic.q .!21' .41~ .-. .C::£ • .2 Q l '3~7 ." 4 "> .3~a .213 .276 .3"6 .&423 .&41q .213 .269 .206 • 2~ .. .2~1 .33~ .3~Q .4(10 .397 0\ " TABLE 5.1.1 (Continued) b = 1.0 a = 0.05 az alf 3 , 9 12 1'1 16 1 1.0 1.5 :>.0 2.5 1.68'1 1.'69 1.331 1.2'12 1.182 1.139 1.108 1.085 1.0~7 1.053 1.033 2.113 1.892 1.760 1.678 1.625 1.588 1.562 1.~'13 1.528 1.517 1.502 2.~~3 2.331 2.2~5 2.129 2.082 2.050 2.026 2.012 2.0na 1.191 1.911 3.000 2.119 2.(,~9 2.~90 2.5111 2.519 2.500 2.'106 2.'175 2.'166 2."52 1.019 1.'192 1.961 2.11'12 1.010 1.'18'1 1.959 2.43'1 2 1.0 1.1If'.9 1.260 1.129 1.ec" 1.~911 1.1176 2.1~7 1.953 1.6U1 2.521 2.326 2.226 1 1.~ 2.0 2.5 2 5 (, 7 8 10 18 20 22 2- 26 28 30 .9811 .9'12 1.003 .998 .993 .990 .981 1.'178 1.'173 1."68 1.465 1.1I~2 1.II~Q 1.1151 1.953 1.9"'1 1.~1I3 1.9'10 1.937 1.9!~ 1.'152 2.'128 2 ... 23 2.418 2.'115 2.1112 2."09 2."01 1.C'l1 .993 .957 .930 .911 .6"6 .893 .8,,5 .852 ./1'12 .835 .Po29 .92'1 .620 .61" .8U • ell 1.'101 1.~62 1.331 1.306 1.291 1.27 7 1.2~6 1.2119 1.237 1."8 1.221 1.~15 1.211 1.207 1.'011 1.201 1.191' 1.71'11 1.142 1.713 1.692 1.675 1.662 1.~Sl 1.635 1.62'1 1.615 1.~r.8 1.603 1.598 1.5 0 '1 1.<;91 1.51'''1 1.'iA6 2.165 2.126 2.098 2.017 2.061 2.0'18 2.03A 2.022 2.011 2.003 1.9"6 1.'190 1.986 1.982 1.919 1.916 1.97'1 .6110 1.0(10 .79'1 .113 .7'13 .7211 .710 .699 .685 .911 .861 .623 .15 7 .6"2 ."1'19 .9911 1.273 1.201 1.154 1.121 1.096 1.076 1.0~3 1.052 1.03'1 1.022 1.0t2 1.0n5 1.560 1.502 1.1159 1.112n 1.1106 1.389 1.375 1.36'1 1.3'17 1.335 1.326 1.319 1.31'1 1.309 1./172 1.810 1.769 1.739 1.718 1.701 1.6n8 1.678 1.662 1.650 1."'11 1.6~'1 1.~2" 1.62 u .616 .669 .e'iJ7 .672 .9P.14 .987 .9Al 1.302 1.29" 1.297 1.617 1.61~ 1."12 3 1,0 1.5 2.0 2.5 1.331 1.59'1 1.875 2.169 1.129 1.390 1.677 1.976 , 1.0 1.'5 2.0 2.5 1.2112 1.452 1.600 1.921 .51'9 1.0'17 .917 .830 .;71 .699 .676 .6'i6 .6'111 .623 .597 .'iA2 .571 .572 .56'1 .!.6'S .5!>3 .729 .6011 .1134 .89/\ .1I6l, .A4'1 .0113 .038 1.251 1.131 1.056 1.006 .926 .'110 .680 .A57 .0.31 .p"o .P2& .971 .9'15 1.1182 1.370 1.301 1.2~6 1.224 1.201 1.1A3 1.168 1.1~7 1.140 1.128 1.111\ 1.111 1.105 1.101 1.0'17 1.093 1.0~1 1.r.F.1I 1.726 1.620 1.S55 1.512 1.4112 1.460 1.1142 1.4?'1 1.'118 1.'102 1.390 1.~81 1. 3"'1 1.'68 1.364 1.360 1.3S7 1. 3S4 1.!~1 5 1.0 1.5 2.0 2.5 1.11'2 6 1.0 .4115 .1140 1.139 .9!17 .611l .537 .521 .496 .467 .458 .451 .436 .823 ".129 .66'1 .583 .557 .480 .1130 ."'~~ .63'1 .879 .82'+ .l,50 1.2112 1.0116 .9f'.1 .784 .756 .734 .71 7 .703 .683 .669 ."!l8 .6 4 4 .1;311 .b~O .62 7 t"~~ .829 .82l, 1.441 1.2113 1.125 1.051 1.0Cl ./\52 .844 .833 .966 .940 .921 .906 .893 .875 .862 .1138 .8< 3 .~20 1.611 1.41'1 1.302 1.233 1.187 1.15'1 1.130 1.112 1.098 1.066 1.069 1.056 1.0'17 1.040 1.03'1 1.029 1.025 1.022 1.019 1.017 1.~ 2.0 2.5 .4'17 .1I9~ .5/\8 .'\23 .51 4 .u87 .993 .861 .771 .709 .664 .632 .f'.07 .573 .!l50 .53'1 .'i07 .501 .1193 .7'19 .720 1.15'1 1.0~2 .953 .'100 .At5 .729 .724 • 71 ~ .767 .743 .663 .836 .786 .753 .735 .71~ .711 r .Q~~ .9 ,9 .9~~ .'I3l, 1.'\42 1.3'14 1.229 1.157 1.110 1.076 1.052 1.0~3 1.018 1.006 .989 .976 .9~6 .948 .'153 .941 1.7'12 1.5117 1.437 1.370 1.326 1.295 1.271 1.25'1 1.2'10 1.'-29 1.212 1.200 1.190 1.1113 1.178 1.173 1.169 1.166 1.163 1.161 .'1~8 .930 .699 .632 .SA:'I .547 1.106 .7"4 .520 .(,95 .~72 1.:'28 1.037 .1125 .766 .725 .910 .A3~ 1.364 1.1611 1.0u7 .~7C .918 .882 .855 1.0;11 1.314 1.200 1.128 1.0nO 1.04l, 1.021 1.00'- .'181 .63'1 .1106 .'376 .456 .f.18 .71\7 .958 .438 .603 .773 .945 .425 .592 .763 .'135 .415 .5AU .755 .928 .1107 .'i77 .749 .92? .401 .60;4 ./\19 .9A8 e 1.0 1.5 2.0 2.5 .773 .676 1.085 .911 1.1FII 1.000 .783 .871 .9.8 1.305 1.111 .908 1.'133 1.237 1.120 1.0116 1.0 1.5 2.0 2.5 1.067 10 1.0 1.5 2.0 2.5 1.~0~ 1.6~1 1.3~~ 7 1.0 1.5 2.0 2.5 , .~qO .74'1 .917 .396 .567 .740 .Q!3 .392 .389 .38l, • ~f4 • .737 .'11u .5": .73u .%7 .731 .905 .'57 .'\10 .665 .821 .354 .5C7 .662 .81A .350 ~ 1&':8 .607 .72? .85'1 .996 .557 .679 .816 .961 .520 .6'17 .788 .935 .491 .(,2:'1 .767 .'116 .4l,9 .604 .7..,0 .900 .1151 .589 .737 .888 .42'1 .567 .717 .870 .405 • 5.51 .703 .856 .392 .540 .l,'12 .8'17 .381 .531 .6A5· .639 .373 • ..,2'1 .(,78 .1133 .367 .673 .628 .362 .514 .669 .82'1 .427 .5u9 .6ill .817 .399 .52(, .660 .798 .379 .510 .6u6 .78'1 .365 .498 .635 .77'1 .35'1 .IIA9 .339 .47l, .f15 .756 .33'1 .471 .611 .751 .329 .u67 ... 07 .74n .325 .11(,4 .60U .7U5 .322 .1161 .~'7 .7~7 .346 .u81 ... 20 .761 .407 .51fo .635 .75A .379 .358 .475 .599 .725 .343 .463 .33'.1153 .579 .7n7 .323 .446 • ..,73 .700 .316 .310 .1135 .5fo3 .691 .306 .u31 .55<;1 .f'.Be .302 .'1211 .298 .420, .553 .6112 .757 .tlul .972 1.259 1.eM. .942 1.371 1.176 1.056 .658 .751 .A60 .980 .588 .foS8 .803 .929 .537 .l,43 .76'1 .892 .498 .610 .734 .866 .469 • 'ill !I .712 .845 .11115 .1"11:'1 1.053 .7113 1.131 .951 .818 1.221 1.031 .906 1.321 1,127 1.006 .6'14 .725 .821 .927 .'\73 .661 .762 .874 .521 .614 .721 .1I3l, .401 .580 .691 .809 .'151 .427 .533 .896 1.1~6 .~72 .S5~ .f'.l,7 .7/\/\ .5~5 .6 5 .830 0 • 6,~q .771 ."92 .(,14 .739 .~Ae .715 .519 .440 .~67 .6'35 .~C)~ .~e~ .~a" .660 .!H5 .c,~2 .742 0\ 00 TABLE 5.1.1 (Continued) b = 1.0 a = 0.05 81 f 82 2 1I II 5 6 7 a 9 10 12 1" 16 18 20 22 2" 26 28 50 .7211 .783 .852 .9:!'1 .625 .687 .763 .848 .550 .620 .701 .7':'1 ... 96 .57t .657 .751 ."56 .535 .62" .721 ."2" .579 ... 67 .565 .657 .3.. 11 ..... 1 .5 .. 2 .647 .326 ."23 .526 .632 .510 .la09 .5tla .1;21 .613 .289 .391 .la98 .f.07 .281 .385 .493 .601 .275 .3AO .~C?1 .399 ."115 .5/10 .6111 .2'11.'1 .3'19 2.~ .865 1.C"3 .'n9 1.166 .'J8t 1.2116 1.055 .270 .376 .484 .593 .266 .372 .4!1 .!.9C .2(,2 .3f.9 .la7/l .'5/11' 1_ 1.0 1.5 2.0 2.5 1.01 .. .852 1.068 .897 1.H7 .9 .. a 1.193 1.006 .710 .5:!,4 ... 80 .541 .612 .690 ... 38 .503 .577 .659 ... 05 .117" .0;0;1 .f.35 .379 .110;0 .530 .616 .35a ."32 .51" .601 .326 .404 ."89 .579 .30 .. .385 .472 .56" .287 .27" .3(,0 .40;0 .544 .26" .351 .4 .. 3 .0;37 .256 .2,.4 .335 .1128 .23'1 .331 .437 .532 .2 .. 9 .339 .432 .5:>7 .23'5 .815 .879 .608 .661 .723 .7q3 .4~ ~ .~23 .520 ."22 .517 16 1.0 1.5 2.0 2.5 1.010 1.0"" .6 0 9 .7 .. 1 .:'192 .343 ."06 .476 .551 .310 .377 .1150 .528 • 287 .356 ... 32 .512 .'F-9 .'''5 .321 ... 01 .2t?~ .Z::H~ .~-96 .~E;t .4('q. .230 .308 .390 ... 73 .22 .. .303 .l4q1 .237 .314 .395 ... 78 .215 .341 .1119 .500 .2"6 .3'0 .409 .219 .0;15 .587 .365 .4'5 .4 rt 3 1.1"4 .""2 .1'1\1 .923 .971 ... 69 ."6f, .!1q .41>3 11.'1 1.0 1.5 2.0 2.5 1. 003 1.0"6 1.077 1.12 .. .'98 .355 ."19 ."88 .2H .33 .. ... 00 .1170 .'''6 .318 .3116 .1158 .2"2 .376 .1149 .'31 .,97 .368 .11 .. 1 .;>;>2 .529 .332 .3Ar, ..... 6 .512 .361 ... 35 .215 .2A3 .356 .4:!O .20'3 .27!! .352 .1126 .2'0'" .274 .3 .. Q .1123 .199 .'71 .3115 .1&20 .t?~!' .323 .:370 ."23 ."81 .289 .339 .245 .300 .:"(,1 .425 .2 .. 1 .2"8 • 3"0 ... 15 .219 .278 .3 4 1 .407 .210 .270 .33 .. .401 .203 .26" .329 .396 .197 .259 .1'11 ... 56 .26" .317 .375 ... 38 .~Z~ .32C .38 P .187 .25C .317 .281 .325 .375 .429 .256 .302 .355 .411 .237 .2116 .340 .397 .2'2 .'273 .3:>1\ .31'>7 .210 .262 .319 .193 .2u8 .~79 .201 .25 .. .:!'12 .372 .:!67 .187 .242 .302 .!63 .275 .3111 .359 .408 .2 .. 9 .291 .33/1 • 389 .':'10 .'74 .322 .37 .. .215 .2(,0 .310 .3(,4 .'03 .:>50 • 301 .355 .193 .241 .29" .3'19 .1115 .235 .2P8 .3113 .179 .229 .2 8 3 .339 .17~ .16~ .2'" .279 .33~ .220 .21", .!32 .22" .'6" .3011 .355 .209 .2'10 .2'l'l .3411 .197 .'39 .335 .187 .230 .278 .329 .179 .223 .21'2 .323 .172 .218 .2(,7 .31& .166 .213 .2f' .314 .161 .208 .2':;9 .311 .191 .,30 .181 .221 .173 .21" .;>59 .306 .16& .HO .20~ .203 .24'1 .2'l7 .150; .199 .293 .155 .1910 .237 .251 .150 .190 .233 .;>711 1 12 1.0 1.5 2.0 1.035 .7~9 .S~2 .659 .733 .507 .600 .~70 .460 ... 0;2 .5~5 .523 .571 .467 .8 .. 2 .597 .642 .69" .752 .~27 .6~0 .579 .6"5 ... 25 .479 .543 .612 .835 .1'61.'1 .905 .9116 .6'12 .7'8 .768 .81 11 .509 .627 .672 .722 .51" .657 .458 .502 .553 .611 .415 ... 62 .516 .57& .:'IAl .430 ... 87 • '1119 1.025 1.060 1.101 .829 ./l51l .8 C1 1 .92S .6Q5 .717 .507 .5113 .5115 .632 .1151 .4(\'1 .53/' .583 .407 .448 .495 .548 .373 .416 .7"2 .562 .616 .655 ,£.98 .520 .3116 .3"'0 .""2 .1198 .82" ./l51 .879 .911 .680 .708 .7"0 .1'75 .577 .607 .641 .679 .SCl .533 .570 .S11 .4"5 ... '" ..... e .3:!,9 .378 .510 .56' .401 ... 37 .479 .525 .367 ... 04 2.~ .'l93 1.017 1.04!! 1.083 .4~'+ .31(, .357 ."05 .4~6 ."7" ."56 211 1.0 1.5 2.0 2.5 .990 1,011 1, 03" 1,%8 .820 .1'76 .7111 .7'1.0 .7"'1 .572 .4'l7 .525 .558 .595 .11"0 .470 .500; .54" .396 ... 28 .4(.5 .506 .:'Ifo2 .3'15 .1&34 ... 77 .3 .... .3(,8 .:no ."~9 .l.6~4 .3117 .311'1 ."36 26 1.0 1.5 2.0 2.5 .987 1.006 .816 .8311 .672 .569 .593 .620 .651 ."93 .519 .5"8 .581 ... 36 ."63 .119" .530 .392 ... 21 ."5" .491 .357 .:'187 .422 .461 .329 .36 0 .3'17 .437 .306 .338 .31'7 .410 .270 .305 .346 .~90 .2 .... .281 .:32'+ .310 ... 90 .'513 .540 .569 .433 .457 .486 .517 .38'3 ... 1 .. .4 .. 4 ... 78 .35" .31\1 .41 ? .~25 .302 .331 ... 0 .. .266 .2911 .3311 .375 .23':1 .2n .312 .354 .219 .;>55 .295 .339 .2n .. .2111 .283 .3;>7 .4B7 .5011 .430 ,:'ISO .;>62 .'91 .325 .362 .235 .;>(,6 .302 .340 .l'l9 .'lIe .2~1+ .47a .507 .298 .325 .357 .392 .'15 .~~3 .386 .409 .436 .467 20 1.0 1.5 2.~ 2.5 22 1.0 1.'5 2.0 1.~9~ .~U4 .P70 ./l9/\ .7~9 .'7~3 .6~6 .5~9 .630 .6611 1.{l~9 .as, 1.056 .8~'1 .7:?1 .749 .'Ht~ .1113 .B3" .IISS .879 .6€-9 .S':'l • 7l'~ .739 .565 .1\11 .830 .61',7 .6 P E- .e~o .708 .731 .563 .503 .60e. .631 28 1.0 1.5 2.0 2.5 1.0C2 1.022 1, ~II7 30 1.0 1.5 2.0 2.5 .982 .99!l 1.0lF. 1.038 .871 .588 .1'.13 .640 .5~5 .6C3 .560 .51q ,uS:? ..... '1 .uF,6 .44A- .:'170; ."04 .436 .5~7 .3';4 ."06 ."F-5 .354 .3 n 7 .1123 .322 .34/1 .37 A .4\1 .366 .39~ .2115 .325 .3~F- .272 .313 .'~6 .3~4 .2Aq .273 .;>6~ .3111 .311 .'22 .262 .\1\7 .177 .213 .2"4 .304 .296 .4~a .5':'7 .3~S .168 .392 .253 .301 .161 .:>os .no .21.17 .290 .242 .226 .2~1.I .~2P. .3P~ .1/l1 .23· .29/\ .177 .231& .29" .3~9 .!:~ .24~ 0\ \0 TABLE 5.1.1 (Continued) b = 1.5 a. = 0.01 81 f 1 2 II :I S , 7 8 , 82 10 12 111 16 18 20 22 2'1 26 28 30 1.000 .9'5 .'e8 .'85 1 1.0 1.S 2.0 2.5 2 1.0 1.5 2.0 2.5 3 1.0 1.5 2.0 2.5 II 1.0 1.5 2.0 2.5 '5 1.0 1.5 2.0 2.5 6 1.0 1.5 2.0 2.5 . .'8' 7 1.0 1.5 2.0 2.5 .9"'1 8 1.0 1.5 2.0 2.5 .982 , 1.0 1.5 2.0 2.5 .'511 10 1.0 1.5 2.0 2.5 .931 .909 .880 .857 .987 .'62 .9'13 .92' .918 .90' .'02 .896 .891 .ee6 .982 .'5'1 .':11 .f'9' .871 .853 .811' .828 .819 .812 .1\06 .801 .7'6 .'8' .,.... .90'1 .880 .857 .821 .796 .777 .7(,3 .752 .7'13 .755 .729 .72'1 .719 .'32 .886 .1l50 .8~0 .7'6 .760 .73.. .715 .700 ~990 .666 .960 .679 .972 .671 .965 .665 .960 .660 .9StS .655 .951 .886 .850 .8'10 .803 .1103 .712 .1'1n .99.2 .710 .961 .683 .939 .663 .922 .6'16 .909 .636 .g99 .62(, .891 .618 .S8'1 .612 • '179 .601, .~02 .87'1 .e10 .592 .1132 .51l2 .823 .57" .e16 .'i67 .811 .501 .806 .'i57 .1102 .16'1 .'114 .733 .948 .106 .669 .1l95 .6'11 .872 .620 .1155 .6~'1 .921 .842 .620 .99" .772 • 95'1 .133 .'120 .701 .8'1'1 .675 .1l72 .63'1 .1139 .605 .1116 .5S'I .198 .51,6 .71''5 .'i55 .77'1 .999 .5'14 .766 • "1 .536 .759 .965 .529 .153 .geo .523 .7 .. " .975 .518 .1 .... .971 .796 .953 .746 .910 .1011 .1176 .675 .6 .. 8 .648 .BU. .606 .792 .992 .576 .7611 .911 .5511 .70;0 .955 .537 .0;23 .7~6 .72!j .9 .. 3 .<133 .513 .716 .925 .5011 .709 .919 .1197 .703 .913 .1191 .69/\ .901' ."86 .69'1 .90" ...., 0 .. TABLE 5.1.1 (Continued) b 1.5 a. = 0.01 al f 6 7 II 9 10 1? 1'1 16 11l 20 22 2_ 26 29 30 .1196 .1l21 • .,113 .760 .811(, .710 .8111 • .,97 .669 .1I0!l .'16' .6311 .776 .':1110 .60(, .7':'1 .':11'1 .'62 .717 .A66 .530 .691 .865 .'06 ."72 • All" .IIA7 .6'17 .1l36 .1173 • toll 6 .1\2(, .1162 .637 .1\17 1.000 .11'" ."'9 .'l11 • 9Q ... .11115 ."23 .1105 .'189 .1139 .61" .900 .9p .. .11"3 ."13 .79(, .92.0 .871 .96':1 .796 .896 .7311 .831\ .':1611 .683 .791 .922 .6111 .75" .1'.119 .6~5 .576 .6':1A .8"1 .':192 .'30 .660 .1196 .633 .78" .':1'+1 .1170 .(,12 .767 .':125 ."'il .5'17 .7'13 .913 ... 36 .'185 .743 .903 ... 23 .575 .7311 .895 ."1" .567 .727 .811'1 ."05 .561 .721 .1183 .39"1 .~93 .723 .8"2 .515~ .~~1 .716 .87" .712 • P.75 .1170 .588 .720 ."'3 .5'\0 ... 07 .'\37 .677 .39" .527 .668 .1112 .383 .519 .661 .805 .375 .512 .(,55 .1100 .31>11 .506 .650 .795 .361 .0;01 .660 .11 .... .566 .702 .1143 .371 .350 ... 72 ... 01 .732 .!4! ."61> .59':' .727 .336 .461 .-:91 .723 .316 .421' .C.""6 ...,,6 It 1.0 1.' 2.0 .ge7 1'1 1.0· 1.' 2.0 2.5 .9Eo2 16 1,0 1.5 2.0 2.5 .9_3 111 1.0 1.5 .929 .999 1~ 2 a2 , 3 1 2.' .853 .935 .839 .909 .99" 2.0 .777 .861 .961+ .763 • 83" .922 2.5 20 1.0 1.5 2.0 2.5 .'318 .979 22 1.0 1.5 • 909 .963 2.0 1.000 .'02 .9"9 2.0 2.5 26 1.0 1.5 2.0 .993 .8 a 6 .939 .9~9 2.'S 28 1.0 1.5 .988 2.0 .O'?l .929 .97' 2.5 30 1.0 1.5 2.0 2.5 .983 .8P6 .922 .9~3 .963 .663 .75'+ .865 .':190 .620 .715 .1130 .'159 .56" .663 .802 .933 .5'1" .657 .780 .913 .o;oe .700 .77.. .866 .971 .6.. 8 .725 .821 .930 .601+ .665 .7115 .1197 .568 .652 .7'15 .871 .537 .625 .732 .8 .. 9 ... 87 .583 ."95 .817 .451 .552 .669 .794 .423 .!l30 .':'50 .777 .401 .512 .(,'1'1 .763 .38" .,,99 ."2" .7'13 ."1'1 .745 .360 ... 80 .607 .738 .436 .524 .6211 .740 .407 .5no .6C8 .722 .3e4 .367 .4(,11 • .,eo ... 97 .352 .457 .571 .689 .3"0 ..... 7 .563 .61'1 .331 .323 .41l2 .'192 .7011 .440 ,434 .'i56 .676 ,551 .670 .325 .421 .">26 ."34 .315 ... 13 .'\19 .628 .306 .40., .513 .299 ."01 .62~ .619 .715 .802 .90'1 .617 .7'+5 .1'82 .(,1'8 .11'10 .1\20 ... 8e .f4~ .731 .828 .81l8 .962 .752 .A14 .e.,o .9711 .6611 .753 .632 .92" .636 .703 .765 .881 .592 .(,62 .7 .. 11 ,1147 .5~5 .523 .626 .718 .820 ... 00 .693 .797 ... 73 .556 ,e55 .76 11 .819 .e72 .7"3 .7.,7 ,513 .579 .661 .755 .4&2 .534 .622 .720 ."23 .501 .593 .695 .~73 .15c.7 .~t&tJ .8l'5 .5.... .6n8 .687 .778 .'152 ..... 2 .941 .'1112 ... 43 .719 .1106 .371 .",\7 .II~3 .626 .685 .757 .841 .39" ... 76 .937 .67'! .736 ."05 ... 76 .6 .. 2 ... '11 .337 ... 30 .53" .""2 .812 .859 .916 .982 .735 .7114. .G42 .911 .€o71 .722 .783 .854 .618 .€o70 .734 .809 .574 .536 .593 .6,:,2 .7 .. 3 .504 ,563 .636 .71'! ."53 .383 .'156 .543 .3 .. 0 .312 .3"9 .495 .59" .302 ,39C .400 .588 .25~ .5?':' .623 .325 ,"03 .503 .602 ,3C3 .4~2 .37e .477 .~37 .3 .. 0 ... 21 .'i13 ... 11 .293 ,595 ,el\3 ."14 ."(12 .565 ,657 .5f3 .~7e .806 .1l48 .1\99 .958 .729 .773 .P25 .665 .710 .76 .. .€o12 .658 .715 .567 ... 15 .7 .. 5 .405 ... 6& .'141 .6?5 .375 ... 40 .,,18 ."04 .350 ."19 .501' .5/19 .331 .403 ... 87 .'177 .315 .390 .476 .5':'8 .291 .371 .u60 .553 .274 .762 .4 .. 5 .502 .572 .652 .281 .36" .145'" .1I2~ ."97 .5 .. 9 ,61" .69n .302 .3110 .8r.6 .529 .5/10 .642 .714 .~~~ .~~~ .eol .839 .M5 .938 .724 .763 .CI0 .865 .660 .700 .606 .648 .699 .759 .'i61 .600; .5?3 .5<,9 ,293 .36" .4"'i .'130 .355 ... 37 .523 .3"'7 .6'10 .30& .374 .454 .538 .272 ."26 ."97 ."7,, .323 .3811 .465 ,'1"1) .2el .6~5 .399 ... 53 ,521 .597 .343 .4C5 ."7,) .721 ."39 .4119 ,'\53 .625 .:!r,6 .1'.'56 .491 .538 .596 .66" .264 .!f.ll '''25 .">13 .486 .528 .581 .643 ... 33 .479 ,'i36 .':'O? .393 ..... 2 .503 .573 .299 .3<'1 , "3" ,512 ,285 .27.. ,3 4 0 ."17 ... 97 , 2,5 2" 1.0 1.5 .I\O~ .796 .832 .873 .921 .719 .755 .7~8 .847 .74') .80':' .655 .692 ,736 .7811 .602 ,640 ,686 .740 .62~ .':'95 .773 .1'.75 .557 .515 .'\9':' .~t,O .644 .701 .610 .669 .516 .361 ."1" .1179 .1Ii\52 .4~6 .5~O ,3~1; .3 0 2 .""0 .5;5 .316 • :a, 7 !") ,446 .523 .~68 .560 • ~~)O ,42" .50" ." 31 .51e .26~ .333 .410 ,491 .~I)' .:!~8 . . If' .255 .326 ,If';5 .4~(, ..... ~ TABLE 5.1.1 (Continued) b = 1.5 a. = 0.05 al f a2 1 2 5 II 5 & 7 8 9 10 12 111 16 18 20 22 2'+ 26 28 30 .988 .962 .9113 .929 .918 .909 .902 .896 .891 .8!!7 .'''13 1 1.0 1.5 2.0 2.5 2 1:0 1.5 2.0 2.5 3 1.0 1.5 2.0 2.5 II 1.0 1.5 2.0 2.5 .989 5 1.0 1.5 2.0 2.5 .923 6 1.0 1.5 2.0 2.5 .967 '.875 7 1.0 1.5 2.0 2.5 .931 8 1.0 1.5 2.0 2.5 .9011 9 1.0 1.5 2.0 2.5 .e1'\2 1.01111 10 1.0 1.5 2.0 2.5 .988 .86'1 .969 .838 .810 .951 .787 .911 .769 .1l79 .923 .857 .807 .769 .938 .739 .887 .715 .846 .9 .. 3 .696 .A12 .9"'1 .9&7 .931 .9011 .882 .86'+ .837 .818 .803 .792 .783 .776 .769 .76" .760 .756 .875 .838 .810 .7,.,7 .769 .7'+1 .720 .705 .6"" .991 .6811 .983 .677 .976 .670 .970 .665 .965 .660 .~57 .96~ .9~7 .807 .756 .95e .716 .892 .68,\ .839 .661 .797 .9'1B .6'11 .76;> .898 .769 .716 .923 .676 .856 .739 .ge3 .715 .9,;2 .696 .9'+6 .667 .920 .6116 .901 .630 .tl87 .618 .876 .608 .1\£,7 .601 .A60 .59'+ .85" .589 .1l"9 .58" .ell5 .580 .685 .1l97 .661 .875 .6111 .858 .610 .631 .588 .812 .572 .796 .5~9 .5119 .777 .541 .770 .535 .764 .995 .52':1 .75'1 .9ge .52" •.71\6 .'120 .7'H .'l83 .644 .82/\ .618 .806 .644 .802 .975 .611 .774 .584 .Q4' .929 .618 .759 .91'1 .584 .72'1 .A87 .557 .7n6 .597 .723 .863 ."i"3 ."93 .A36 .985 .7~1 .8F-6 .5",5 .6<,8 .81'1 .964 .7~~ .96f .~II1 .597 .788 .966 .566 .761 .963 .5113 .7111 • ':1115 .526 .726 .931 .513 .7t5 .920 .502 .706 .912 ."911 .696 .905 .'+87 .692 .899 ... 81 .687 .89 .. .476 .682 .890 ."72 .f7S .563 .732 .912 .~30 .70" ./\86 .506 .68" .867 ... 89 .669 .1'\53 .475 .6'17 .8 .. 2' ... 65 ..... 7 .A33 ."56 .6.. 0 .1l26 ..... 9 .633 .A20 ..... 3 .628 .A15 ."'/\ .623 ."33 .619 .e07 .995 .501 .657 .e22 .990 .477 .637 .P03 .972 .459 .<';>1 .7AA .958 .445 ... 33 .,\99 .766 .'139 .4211 .591 .761 .932 .417 .5115 .755 .926 ... 11 .57'1 .750 .921 ."06 .575 .7"5 .917 ."01 .609 .777 .9 .. 7 .'177 .619 .769 .922 .452 .597 .749 .90'+ ... 33 ."illl .73'+ .1\90 .'119 .51'.9 .723 • 879 ... 08 .'1'19 .391 .3B" .37" .54 4 .~311 .53\6 .71~ .39A .551 .706 .BE>3 .695 .852 .690 .A70 .700 • 857 .37" .&30 .667 .535 .6A7 .1\ .. 8 .512 .6"9 .796 .9"8 .All .99'1 .8"P .~~6 .~71 .7~2 .913 .e".. ..... N TABLE 5.1.1 (Continued) b = 1.5 ex = 0.05 al f a2 2 3 12 1.0 1.5 2.0 2,.5 .962 .837 .922 1/1 1.0 1.5 2.0 2.5 .9"3 1 16 1.0 1.5 2.0 2.5 18 1.0 1.~ .929 .9l)7 2.0 2.5 20 1.0 1.5 2.0 2.5 .909 .95" 22 1.0 1.5 2.0 2.'5 • 902 .9'12 .9117 2'1 1. a 1,5 2.0 2.5 .896 .933 16 18 20 22 211 26 .11'11 .558 .685 .!l15 ."1'1 .535 .66'1 .796 .395 .518 .6'19 .71\2 .379 .51)5 .637 .771 .367 ... q!\ ... 27 .3'\7 ... /17 .620 .75'1 .3"9 .'11'0 .613 .859 .1177 .590 .713 .8"2 .3'13 .117" .606 .7" 3 .'152 .5 .. 7 .652 .763 ."1" .51" .366 ." 72 .51lS .700 .337 .""7 .327 .319 ,432 .4~1 .572 .6·9 .f)~2 .~~4 .735 .387 ."90 .601 .715 .3"i0 .6'16 .1\0 3 .'177 .5<,9 .672 .7nl .679 ... 72 ... 6'> .5"~ .He .'189 .'165 .651 .7'1'1 .t+t!\9 ."33 .'115 .605 .702 .395 .'190 .57'1 .673 .36& .'155 .552 .652 .3.. 5 .'136 .'135 .f>37 .3211 ."22 .5;>2 .6;>5 .315 .'111 .30'1 ."C'1 .'103 .608 .295 .39" .497 .601 .21:18 .387 .626 .721 .'1 .. 5 .513 .'119 .'56'1 .65" .328 .1108 ... '1'1 .5116 .297 • 3e 1 ."71 .2e6 .372 .5 0 .350 ."28 .513 ... 02 .311 .3 0 3 .61".697 .379 ."54 .536 .1'02'1 .• 574 • ~6'+ .502 .558 .621 .691 .'165 .'1;>2 .'1111l .660 .'133 .'193 .51'01 .'1011 .'169 .539 .615 .367 .'133 .506 .511'1 .337 .'106 .1182 .562 .315 .3F1& ."6'1 .5 .. 5 .2'17 .370 .'I"iO .533 .283 ... 911 .5'13 .600 ... 62 .'156 .'107 .'1," .U78 .'<>"6 .~"7 .. 3~1 .5~1I .327 .38/l ... 56 .:'I~" • ~C;(I ... 37 .630 .6r'l .398 ... 53 .'11'" .5/!3 .357 ."1F, ."1\1 .75" .5'11 .58'1 .6"2 .70:> .~51 .~211 .670 .710 .7511 .1'02 .5'111 .635 .681 .732 .535 .577 .625 ... 7" .'187 .531 .582 .637 ..... 9 ... 9'1 .5 .. 7 .605 .'117 .'1"'5 .519 .578 .391 ..... 0 .1196 .557 .3'19 .'101 ."60 .52" ."11 .3'13 7 8 .7111 .1I30 .933 .667 .761 .869 .986 .~10 .70" ./121 .9'11 .566 .668 .78'1 .907 .530 .636 .75'\ .Ml .6"6 .72'1 .C1" .913 .588 .670 .76'1 .867 .5.. 3 .629 .726 .1l31 .50G .705 .769 .8 .. 2 .922 .630 .6')7 .773 .858 .5 7 2 .6'12 .722 .610 .526 .599 .682 .773 2 .,. .. 5 .90" .971 .6':1'1 .7"9 .812 .881 .618 .676 .7"2 .Fl15 .55'1 .620 ... 8'1 .76F, .513 .576 .6"8 .727 .763 .830 .6a'l .733 .Te8 .8'19 .608 .659 .717 .782 • 5'19 .603 .663 .731 .601 .6'16 .915 .677 .720 ,i(-,9 .11:>3 .759 .807 .849 .8"5 .l!1)3 .6FII .7 0 .1\1'2 .9 .. 0 .~7, 1_ (, .9~3 .918 .969 12 5 .818 .f!.1'9 .970 .776 .817 .e~" , 10 " .720 .795 .881 .')76 .#;97 ."qS .1175 .'1 .. 1 .5~1 .611 .7~2 .5~7 0 .673 .~~S .4n~ .F.22 .&lee .7"~ .'il? 74~ .1416 .0,33 .6~h .';~2 .2P2 .276 .37 0 ."€2 .'187 .3,c:.? ."~" .~~~ .44 .550 .'5'4'+ .53Q .'\3'1 .439 .'122 .271 .3 .. 8 .'130 .'11" .262 .3" 0 ."23 .507 .25 .. .333 .'117 .502 .2"7 .3;>7 .'111 ."97 .;>"2 .322 .'-07 ."93 .260 .3;>1J ... 03 .250 .;>2'1 .3~~ .~79 • ~11 .U~A .4.0.8 .47Q ... 72 .2"2 .313 • ~,89 ... 67 • 2 3£. • ~t, 7 .'I2~ .:>71 .33'1 ... 12 .""'2 ."'\7 .319 .373 ... 35 .501 .;>95 .352 ."16 .41\3 .277 .;>62 .323 .;>50 .31;> .231 .291> .2~t;, .290 .~~~ ... 0 1 .46'1 .~89 .~80 .~65 .3~r: ."59 ."50 .2 .. 0 .303 .372 ..... 3 ."37 ."3;> .355 .";>7 .312 .361 ... 17 .'177 .288 .33'1 .269 .3?3 .2~t.J .231 .2 P Q .209 .27C .~q7 .3~2 ... 45 .'13'1 .'118 .223 .2"'2 .3"& .'111 .27~ .'159 .2"2 .2"1J .360 .'125 .21£' .309 ..... 3 .501 .3 .. n .'106 ... 02 .247 .:>911 .;>35 .;>87 • 3~ 4 • ~'I" .21& .270 .32e .369 .20P .263 .323 .38" .;>02 .!.A2 .263 .311 .31;6 .u .. .209 .25'1 .202 .253 .1'15 .;>"7 .!O3 .3&0 .3~6 .~'I8 .32:: .~q~ .529 .567 .511 .659 ."81 .521 .567 .617 ... 43 .532 .56'1 .557 .38" ."29 ... 79 .535 28 1.0 1.5 2.0 2.5 .887 .917 .951 .f>('0 .693 .730 .769 .511'1 .618 .656 .699 .52'1 .559 .59'1 .6'13 .'176 .513 .55" ... 00 .1131' .1175 .519 .'1 .. 6 .'Ir6 .'1 .. 5 .'1119 .539 .379 .'119 .'1&0; .51" .337 .380 .'129 ... 82 .30(, .351 .'102 ."57 .2112 .3;>9 .981; .71-0 .7°2 .826 .86'1 ... 38 .u,:?4 .'112 .403 .22'1 .277 .335 .396 30 1.0 1.'5 2.0 2.'5 .8113 .912 .14;> .975 .756 .7rS .817 .8·2 ... 57 .687 .51'0 .611 .6 .. 1; .6P'I .5;>0 .55:? .51.'9 .629 .'172 .5C6 .5'1'1 .585 .'1B .1I6F1 .'\07 .551 .'I~1 .37" ."11 .115 4 .500 .33;> .371 .416 .'165 .301 .3"2 .3119 .'139 .n6 .319 .3"8 .'120 .;>57 .302 .3·.2 .'106 .2'12 .288 .229 .;>77 .329 .3115 .;>18 .267 .3;>1 .377 ."76 .523 .301 .~2~ .537 .SSf .589 .£,2& .1'067 .713 .756 .306 ."6~ .519 .;>63 .665 .701 .7111 .7/1'1 .'I~7 .7~S .269 .3'\7 .450 .277 .76" .7 e1 9 .6:'17 .878 .7~O .60! .739 .3"" ."'16 .891 .'12'1 .9(,1 1.000 ."s" .5n3 ... 91 ..... " .591 26 1.0 1.5 2.0 2.5 ... e'l .312 .425 .5 4 2 30 .332 .50;<, ."~;> .615 • 26 .337 • ~'90 ."~ 70 •.'\V? .39'1 .~c)2 .~7C .3~~ c .3t)j:l .365 .257 .3"7 • 144 1 .!~2 .211J .~3'1 .2~'! .318 .3PO ...... W TABLE 5.1.1 (Continued) b = 2.0 a = 0.01 al f 1 2 8 5 6 7 " 1.717 1.628 1.55" 1."93 1 ..... 1 t.39A II , a2 10 12 1" 16 18 20 22 2- 26 28 30 1 1.0 1.5 2.0 2.5 1.'28 2."11 2.891 3.385 1.821 1.3~2 1.331 1.281 1.2_2 1.212 1.1~7 1.161 2.2'5 2.107 2.0'6 2.022 1.961 1.'11 1.A6' 1.83_ 1.S0" 1.756 1.71' 1.~'0 1.641 1.6"8 2.17.. 2.662 2.511 2.491 ~ ... 37 2.361 ~.~"1 2.312 2.283 '-.231 2.2~1 2.11" 2.1~2 ~.135 3.256 3.1"2 3.050 2.916 2.'11 2.868 2.82A 2.1 a 5 2.167 2.722 2.68~ 2.661 2.~40 2.62" 1.1~0 1.136 1.12.. l.ll" 1.105 1.~33 1.620 1.609 1.5~' 1.~91 ~.120 2.10' ~.097 2.0pP 2.0fl 2.~10 2~598 2.588 2.~.0 2.~72 2 1.0 1.5 2.0 2.5 1.821 2.265 2.715 3.163 1.695 2.126 2.566 3.013 1.273 1.239 1.210 1.165 1.131 1.105 1.0p" 1.06~ 1.106 1.6~5 1.6"9 1.608 1.578 1.~55 1.5'8 1.52" 2.153 2.12" 2.100 2.063 2.036 2.016 2.000 1.~88 2.601 2.560 2.55R '.524 2.500 2."01 2.467 2.455 1.054 1.043 1.03.. 1.0~o 1.0~O 1.513 1.~0" 1.496 1.4q~ 1.484 1.978 1.970 1.~~3 1.9~7 1.9~2 2.44E 2.438 2.432 2.~26 2.4~1 :5 1.0 1.5 2.0 2.5 1.711 1.535 1.415 2.116 1. 9 70 1.8~5 2.523 2.368 2.250 2.93~ 2.773 2.656 1.385 1.765 2.162 2.510 1.313 1.695 2.095 2.505 " 1.0 1.5 2.0 2.5 1.628 1.986 2.351 2.722 1.296 1.63" 1.991 2.360 1.224 1.166 1.119 1.080 1.565 1.511 1."~6 1.434 1.926 1.e7~ 1.B36 1.A05 2.291 2.250 2.213 2.1A3 5 1.0 1.554 1.422 1.313 1.22" 1.152 1.094 1.041 1.875 1.72A 1.613 1.525 1.~~6 1.~03 1.360 2.20 4 2.0~8 1.9~2 1.~46 1.101 1.731 1.692 2.539 2.318 2.263 2.179 2.111 2.069 2.033 1.009 .971 .950 .908 .818 .85" .637 .~22 .8U .801 .19,5 .1e6 .7eo 1.326 1.298 1."5 1.239 1.'13 1.19" 1.178 1.166 1.156 1.141 1.140 1.134 1.129 1.661 1.635 1.61" 1.~e2 1.55A 1.540 1.5'6 1.514 1.505 1.4~1 1.490 1.~p~ 1.4eO 2.003 1.979 1.959 1.929 1.906 1.AA9 1.B75 1.~65 1.956 1.B48 1.8"2 1.836 1.831 6 1.0 1.5 2.0 2.5 1.493 1.362 1.25" 1.iA' 1.636 1.5'3 2.0~0 1.925 1.810 2.384 2.224 2.109 1.23~ 1.2~1 1.5~8 1.512 1.84~ 1.824 7 1.0 1.5 2.0 2.5 1.441 1.704 1.916 2.253 e 1.0 1.5 2.0 2.5 1.3 a 8 1.273 1.161 1.080 1.009 .950 .902 .AE' .820 .aoo .75f, .724 .699 1.639 1.496 1.30" 1.2 9 7 1.228 1.173 1.129 1.093 1.06" 1.040 1.000 .916 .955 1.·e7 1.735 1.621 1.~35 1.4(9 1.417 1.377 1.345 1.318 1.296 1.2&3 1.23A 1.'19 2.141 1.ge3 1.869 1.784 1.721 1.613 1.635 1.~04 1.5aO 1.559 1.528 1.505 1.4A1 , 1.0 1.5 2.0 2.'5 1.362 1.239 1.135 1.048 .971 .918 .869 .1\28 .79" .166 .121 .681l .662 .643 .627 .615 .604 .596 .5.' .582 1.~~2 1.442 1.331 1.24~ 1.174 1.119 1.07" 1.039 1.0~9 .91\4 .946 .918 .891 .pal .A61 .P57 .348 .8~0 .83'1 .829 1.~11 1.~&1 1.541 1.4~1 1.3 0 4 1.342 1.301 1.~69 1.242 1.220 1.185 1.160 1.1~1 1."6 1.114 1.10~ 1.0'S 1.~89 1.0~3 1.:7. 2.~45 1.~S9 1.774 1.&90 1.~26 1.511 1.538 1.50~ 1. 4 03 1.4~2 1.430 1.40f, 1.,p8 1.37" 1.~63 1.35" ~.346 1.'3~ 1.3~" 1.~29 10 1, 0 1.5 2.0 2.5 1.331 1.210 1.1r1 1.0?1 .9!OO .1\90 .e"l .AOO .766 .731 .691 .657 .631 .611 .595 .582 .571 .~62 .~5S .'54~ 1.53" 1.396 1.2~6 1.1'9 1.12'3 1.073 1.028 .991 .9<;1 .93'" .e97 .C6Q .1\~1 .1\'0 .A17 .806 .1 a 7 .769 .783 .777 1.146 1.~97 1.404 1.397 1.330 1.'-1~ 1.236 1.203 1.176 1.153 1.118 1.093 1.073 1.05P 1.0"6 1.G36 1.0~5 1.021 1.Gl~ 1.0GQ 1.9E2 1.801 1.G a 3 1.600 1.54~ 1.49~ 1.455 t.42" 1.399 1.370 1.345 1.321 1.303 1.209 1.271 1.26A 1.260 1.253 1.24A 1.2"3 1.~ 2.0 2.5 1.495 1.838 2.195 2.560 1.313 1.506 1.882 2.094 1.585 1."95 1."22 1.362 1.313 2.011 1.9~0 1.04A 1.790 1.7.... 2."49 2.358 2.2~8 2.232 2.188 2.09" 2.fO" 2.136 2.683 2.6"1 1.385 1.123 2.018 2.44" 1.206 1.4'" 1.706 1.979 1.254 1.6"0 2.0"3 2 ... 51 1.206 1.161 1.135 1.107 1.065 1.596 1.~61 1.532 1.50A 1.411 2.003 1.971 1.945 1.923 1.890 2.419 2.3A9 2.364 2.3 .... 2.'13 -1.166 1.094 1.036 .989 1."35 1.~66 1.312 1.2E9 1.72" 1.659 1.609 1.569 2.025 1.963 1.915 1.818 .950 1.048 1."06 1.780 2.160 .'318 1.033 1."45 1.866 2.291 1.010 .9 9 1 .'H7 .965 .9~5 .947 .940 .«:lI!4 1.4'5 1.4n9 1.'91 1.386 1.378 1.371 1.3~5 1.359 1.q48 1.8'3 1.A22 l.el2 1.80" 1.790 1.7 9 ) 1.7e7 2.'73 2.260 2.'4' 2.239 2.232 2.225 2.2~C 2.215 1.021 .980 .949 .926 .908 .99" .883 .1113 .865 .a~e .8S3 . 1.363 1.347 1.321 1.302 1.2~6 1.,7~ 1.26~ 1.256 1.24p 1.24~ 1.237 1.159 1.126 1.102 1.~A" 1.670 1.659 1.6~9 1.641 1.~35 1.~29 1.~24 2.1"0 2.109 2.067 2.070 2.0~6 2.045 2.036 2.029 2.322 2.017 2.012 .890 .848 .817 .79&4 .776 .761 .1 .. 9 .740 .131 .125 .119 1.183 1.1'" 1.121 1.101 1.0A5 1.n73 1.063 1.05~ 1.041 1.041 1.036 1."91 1.'158 1."3" 1.416 1."~2 1.~90 1.3~1 1.373 1.366 1.3~~ 1.3~~ 1.005 1.114 1.151 1.734 1.1'0 1.709 1.700 1.692 1.6B6 1.6~n 1.67E 1.119 1.047 .989 .9 .. 1 .902 .869 .841 .198 .167 .143 .12" .109 .697 .681 .679 .6 7 2 .f6€l 1.360 1.291 1.23~ 1.193 1.158 1.130 1.106 1.069 1.0"2 1.022 1.006 .994 .983 .975 .9E7 .c?t;1 .q~6 1.f21 1.5~6 1.505 1.465 1.43" 1."08 1.3b6 1.353 1.328 1.310 1.2~5 1.28 4 1.274 1.266 1.,!~ 1.2~4 1.2~9 1.A95 1.832 1.784 1.741 1.111 1.693 1.612 1.~41 1.618 1.601 1.sn1 1.516 1.561 1.~~3 1.553 1.547 1.5'" .6 A O .~:l.9 1.~04 1.~73 .~65 .653 .6 .. 3 .634 .627 .e-21 .926 .915 .907 .A99 .e9~ .e~9 1.193 1.1A3 1.175 1.168 1.162 1.151 1.462 1.452 1."45 1.436 1.~33 1.428 ...., ~ TABLE 5.1.1 (Continued) b = 2.0 a = 0.01 al f 1 :2 3 II 5 6 7 8 , az 1'1 16 18 20 12 1.0 1.5 2.0 2.5 1.165 1.065 .'80 .'08 .8'18 .7'8 .756 .721 .6'1 .6'13 .608 1.'156 1.323 1.21'1 1.127 1.057 1.000 .'5'1 .916 .8e5 .856 .816 .788 1.63'1 1.4''1 1.3A2 1.2'5 1.227 1.174 1.131 1.0'16 1.0~8 1.0~5 1.009 .982 1.~2? 1.~75 1.5~1 1.476 1.'110 1.360 1.320 1.267 1.2"'1 1.''10 '.~06 1.161 .!lSO .765 .9"'2 .5,\' .542 .734 1'+ 1.0 1.5 ::>.0 1.2'12 1.1:'11 1.0:'13 .9~9 .878 .817 .71:.7 .72'1 .6118 .657 .f.08 .571 .'1'12 .5'0 .0;02 .~"8 .~76 .4~6 1.:395 1.266 1.1<,0 1.073 1.003 .9'15 .898 .65'" .826 .799 .757 .726 .702 .61''1 .~69 .6'\8 .6148 .64Q 1."53 1.~1! 1.3r'l 1.217 1.1"9 1.09~ 1.050 1.Cl'1 .9a5 .%1 .923 .696 • A75 .A<;9 .D~E) .635 .D;;t6 .1'19 1.722 1.573 1.'I~0 1.37'1 1.307 1.255 1.21'1 1.181 1.15'1 1.132 1.097 1.072 1.053 1.0'7 1.025 1.016 1.007 1.000 2.~ III 12 1.2~1 1.212 1.10~ 1.010 .926 .85'1 .79'1 .743 .(,99 .61'2 .631 .580 1.!>'IP 1.22~ 1.116 1.032 .%1 .9C3 .e·55 .1115 .71\1 .753 .709 1.49/l 1.35' 1.2"'1 1.1'\7 1.087 I.O~:> .987 .'150 .9,0 .89'\ •• '16 1.638 1.'192 1.3ro 1.293 i.226 1.173 1.131 1.0'17 1.0<,9 1.0"6 '.010 .5'12 .677 .1127 .98'1 .513 .€-52 ./I/lE. .9(,'1 18 1.0 1.5 2.0 2.'5 1.1/17 1.3C9 1.'137 1.'570 .'IA9 20 1.0 1.5 2.0 2.5 1.167 1.068 .977 .89'1 .6:>2 1.277 1.159 1.0~9 .97'1 ."03 1,~93 1.263 1.1~6 1.070 .99'1 1.«13 1.374 1.26Q 1.111 1.106 22 1.0 1.5 2.0 2.5 .761 .8'1'1 .'42 .611 .717 .8'11 .97(, .559 .672 .1'01 .'3'1 .520 .638 .771 .91:> .7'19 .1''12 .6,7 .717 .825 .944 .595 .688 .798 .919 .5'12 .1'.'11 .756 .1\81 .502 .60t. .72t. .853 .'171 .'11'0 .702 .1\32 .615 .6'1'1 .7'10 .6'17 .562 .664 .762 .871 .529 ."'16 .719 .fl32 .'188 .580 .688 .803 .'156 .6~'I .665 1.05~ .709 .794 .895 1.C09 1.150 1.250 1.3'56 1.466 1.054 .965 .883 .611 .7'19 1.136 1.0~8 .95'1 .~!!2 .822 1.22'1 1.1''1 1.038 .966 .910 1.329 1.220 1.13'1 1.06'1 1.009 .697 .772 .862 .96'1. .653 .730 2_ 1.0 1.5 2.0 2.5 1.136 1.228 1.325 1.426 1.043 .C:;~5 .873 .eOl 1.117 1.0'0 .937 .A(,5 1,2Cl 1.0'18 1.012 .941 1.292 1.18'1 1.098 1.G28 .7'10 .805 .A83 .972 .687 .75'1 .A34 .926 .<''13 .712 .79'1 .1188 .675 .7(.1 .6.,7 .571 .6'15 .733 .631 .517 .595 .<'1111 .791 26 1.0 1.5 2.0 2.5 1.12'1 1.2C9 1.298 1.391 1.034 .~47 .865 1.101 1.006 .923 1.171 1.075 .9.,0 1.260 1.15'1 1.068 .793 .1\51 .919 .998 .731 .790 .660 .9'10 .679 .739 .Al1 .119'1 .(,3'1 .(,96 .771 .1156 .5"'6 .61'>0 .736 .82'1 .562 .626 .708 .797 28 1.0 1.5 2.0 2.5 1.11'1 1.1'12 1.'75 1.31'2 1.026 1.087 t.157 1.233 .940 .658 .9"3 .911 1.0'16 .972 1.128 1.0112 .786 .639 .901 .972 .7:>5 .778 .8'11 .91'1 .672 .727 .792 .A67 ."'27 .51\8 .646 .716 .795 30 1.0 1.'5 2.0 2.'5 1.105 1.177 1.255 1.335 1.020 .934 .1\53 1.07'5 .9113 .900 1.B9 1.0'10 .9% 1.210 1.lC6 1.020 .780 .1l29 .8115 .9'19 .71'1 .7611 .625 .A91 .666 .716 .775 .8'13 .51\2 .6'5 .6')8 .771 < .70;~ .1157 .973 .1123 .9:>7 .1'.83 .751 .~~A .1'.21 .672 .733 .110'1 .9~6 1.1~3 1.1~8 16 1.0 1.5 2.0 2.5 1.06'1 .9"1 .9/\6 .637 .776 .72'1 .hAO .6'13 1.l/1A 1.0~6 1.0nO .Q?9 .67/\ .621 .760 • 7~6 1.303 1.1"5 1.109 1.039 .962 .936 .899 .867 1.'127 1.316 1.229 1.1"'1 1.107 1.06'1 1.029 1.000 .7~8 .~12 .'189 ."'33 .7"Q .9"9 .'1"5 .5'" .732 .87~ 22 26 26 3C .52' .517 .508 .723 .713 .705 .'13" .923 .915 .9~S, 1.136 1.127 1.119 1.11' .500 .l.9'l .9C;> .4'4 .471 .<,111 .77'5 .936 .4'16 ... 77 .718 .11(,3 2" .9«; .. .~O7 .9iot Q .ld3 .!J96 .-:£7 .~2" .~~~ .76'1 .926 .70;5 .916 .7~~ .741 .!?:. .B(' .'11 .%5 .1\99 .'130 ."17 .~17 .-;C.3 c~9r:! .S~7 .697 ."07 .0;'14 .689 .3'17 .!.64 .707 .1153 • (,83 .'531 .677 .P4'1 ,°37 •P~1 .~~., .396 .516 .6'19 .733 .385 .37~ .367 .5~9 .5Cl .l.6q~ .641 .776 .63!4 ,7;Q .62 0 .76 .. .3~7 .349 .471 .593 .717 .465 .816 .f'!)3 .782 .'1'0 • 5 ~1 .645 .7(,5 .'110 .0;1'1 ."'30 .751 .393 .500 .618 .7'10 .379 .'189 .606 .731 .367 .'179 .600 .72'1 .'176 .558 .656 .761 .4'13 .0;30 .631 .739 .4t7 .5n8 .611 .722 .~96 .490 .0;36 .708 .379 .47(, .696 .36'1 .'11'>'1 .574 .687 .352 .'154 .565 .679 .506 .578 .f.62 .756 .'166 .5'10 .628 .72'5 .'133 .'Itl .F.03 .702 .4n7 .4.ne .367 .455 .55'1 .659 .352 .6A4 .385 .'170 .0;67 .1',70 .6 4 9 .555 .615 .686 .766 .500 .56'1 .f.40 .72'5 .'158 .525 .605 .694 .42'1 .'195 .579 .1'.70 .317 .'171 .50;6 .6'12 .~75 ... 53 .'\112 .1',37 .357 .'137 .529 .626 .5118 .603 .6(,6 .7113 .49'1 .551 .1.21 .f.9'1 .'151 .512 .585 .667 .'117 .'11\1 .556 .1','13 .390 .'1<;7 .537 .6''1 .'67 .1138 .')20 .t.09 .349 .422 .506 .597 .5P3 .'=, .. '''~1 .(,2 7 .443 .61'4 .f (,'1 .4!A .633 .60; .'110 .5:>9 .656 .792 .~53 .";"- .~56 .'126 .0;'12 • .;70 .446 .5'1'1 .~'3 1.1~' 1.1~1 .~8'1 .~!7 .712 .3"2 .333 .4 .. ; .673 .'139 .'\'52 .667 .3"0 ... 33 .535 .6"1 .329 .'12'1 .528 .63'1 .320 • "1 7 .521 .628 .3 4 2 .'12'5 .518 .616 .329 .414 .509 .31 0 .405 .309 .5~1 .60£1 .6Cl .'195 .59'1 .333 .'109 .'195 .587 .320 .398 .'Ie6 .0;78 .30'1 .389 .'178 .571 .443 .5 4 '+ .~5~ .!9! .29" .3~1 ... 71 • c.~~ --.J VI TABLE 5.1.1 (Continued) b = 2.0 a = 0.05 al f 1 2 22 2'1 1.560 1.'167 1.3" 1.3'10 1.2" \.~5~ 1.2~0 \.250 \.166 1.137 1.115 1.0~8 1.084 2.011 1.9'0 1.850 1.797 1.755 1.721 1.6~3 1.~70 1.63~ 1.609 1.~A9 1.573 1.~61 2.472 2.382 2.315 2.263 2.2~3 2.1°1 2.1~5 2.143 2.110 2.086 2.a~8 2.0~4 2.0 4 3 2.939 2.6~0 2.7115 2.735 2.~96 2.~~6 2.641 2.621 2.590 2.568 2.551 2.5~8 2."2e 1.073 1.551 2.034 2.S20 1.06" 1.543 2.02 7 2.513 II 5 1.832 1.681 2.2~~ 2.134 2.7i2 2.~95 3.233 3.C62 2 1.0 1.5 2.0 2.5 1.661 1.524 1.404 2.0~2 1.~15 1.795 2.4CJl 2.320 2.2Cl 2.906 2.733 2.617 3 1." 1.5 2.0 2.5 1.Sf1 1.404 1.2115 1.196 1.1~9 1.903 1.73~ 1.620 1.535 1.473 2.257 2.08~ 1.972 1.691 1.1133 2.617 2.446 2.334 2.256 2.1CJ9 1.0 1.5 2.ll 2.5 a2 20 3 1 1.0 1.5 2.0 2,5 'I , 6 7 8 10 12 1'1 16 111 26 28 30 1.056 1.049 1.04. 1.536 1.531 1.';26 2.021 2.016 2.~12 2.508 2.~C! 2.'CO 1.314 1.245 1.193 1.151 1.116 1.0~1 1.069 1.035 1.011 .<]')2 .976 .967 .957 .<]50 .<]43 .'n7 .Q!3 1.708 1.644 1.596 1.559 1.529 1.5n6 1.407 1."57 1.43~ 1.41') 1.406 1.~95 1.3~7 1.379 1.373 1.36~ 1,3~3 2.118 2.0~6 2.014 1.979 1.q~2 1.9~1 1.913 t,P65 1.664 1.~46 1.6~6 1.~26 1.n17 1.Al0 1.P.04 1.7'19 1.7'15 2.S37 2.481 2.43CJ 2.406 2.'AO 2.3~9 2.342 2.315 2.295 2.279 2.2~7 2.257 2.249 2.242 2.236 2.231 2.~26 1.077 1.426 1.789 2.157 1.036 1.004 .978 .956 .n3 .e99 .Illli .A66 .A55 .845 .637 .~~1 .P.o2 c .::;~~ 1.390 1.'61 1.3'8 1.319 1.21l9 1.'67 1.?50 1.237 1.?26 1.217 1.210 1.?C4 1.19" 1.1<;4 1.755 1.72n 1.7r6 1.66n 1.660 :.639 1.623 1.6tO 1.~00 1.S91 1.584 1.:78 1.573 1.'68 2.125 2.0~~ 2.077 2.0;0 2.033 2.012 1.9'17 1.9~4 1.974 1.966 1.9~9 1.953 1.94P 1.9~~ 1.467 1.314 1.1°6 1.108 1.040 .951\ .947 .Q14 .8A6 .666 .832 .607 .71\9 .774 .762 .752 .744 .738 1.7f4 1.600 1.4~3 1.399 1.337 1.289 1.253 1.2~3 1.2~0 \.lAO 1.150 1.127 1.110 1.097 1.01\6 1.071 1.070 1.G6~ 2.r70 1.903 1.7~P, 1.708 1.6h9 1.604 1.S70 1.542 1.520 1.502 1.473 1.452 1.436 1.423 1.412 1.4r4 1.~"7 1.391 2.~84 2.215 2.102 2.024 1.967 1.925 1.892 1.A~5 1.8'14 1.826 1.799 1.77A !.7~2 1.7~O 1.740 1.731 1.72 4 1.718 .73~ .727 1.G"~ 1.~53 1.~~~ 1.~~1 1.71~ 1.70Q 5 1.0 1.5 2.0 2.5 1.396 1.654 1.923 2.1 9 9 1.245 1. 4 93 1.75/\ 2.'31 1.123 1.040 .972 .919 .878 .844 .617 .7~5 .760 .735 .715 .700 .6l18 .E.76 .6 7 0 .663 .&57 .6~2 1.317 1.2~2 1.229 1.1~1 1.143 1.114 1.0~9 1.06~ 1.03~ 1.016 .998 • Qll.5 .'17l4 .Q65 4951 .~50 • '1"t~, .Q40 1.6 4 3 1.5~2 1.50.'- 1.457 1.422 1.394 1.371 1.353 1.3l3 1.302 1.2A5 1.'72 1.262 1.253 1.246 1.24C 1.23 4 1.230 1.918 1.~39 1.7112 1.739 1.705 1.~79 1.6~7 1.639 1.611 1.590 1.574 1.'i~2 1.<;'i1 1.543 1.536 1.530 1.'52 4 1.C:2a 6 1.0 1.5 2.0 2.5 1.340 1.568 1.1\0' 2.C51 1.193 1.409 1.64'1.e"'4 1.077 ,968 .919 .665 .623 .789 .7F.l .736 .702 .676 .656 .641 .1;29 .08 .610 .£.03 .597 .591 1.2"3 1.207 1.144 1.095 1.056 1.026 1.001 .981 .949 .926 .906 ,894 .1\~3 .874 .666 .1\59 ,854 .P49 1.527 1.445 1.3/15 1.339 1.303 1.'75 1.252 1.233 1.203 1.181 1.164 1.1'11 1.1"0 1.132 1.12 4 1.118 1.11~ 1.1C~ 1.771 1.6CJl 1.633 1.589 1.555 1.~2'" 1.5~6 1.4,.,1\ 1.459 1.43/1 1.422 1.409 1.399 1.391 1.31\3 1.377 1.372 1.!67 7 1.0 1.5 2.0 2.5 1.29S 1.151 1.036 1.711 1.~49 1.4~4 1.930 1.765 1.6~1 8 1.0 1.5 2.0 2.5 1.259 1.441 1.633 1.831 1.11/1 1.00 4 .91 4 .8 4 4 .789 .745 .709 .6.0 .656 .616 .591 • ";70 .5'13 .540 .530 .521 .513 1.2A7 1.172 1.0/\5 1.019 .969 .929 .AQ1 .871 .850 .017 .79'- .774 .759 .748 .738 .730 .723 1.47? 1.3~7 1.274 1.?12 1.164 1.127 1.0~~ 1.074 1.054 1.023 1.001 .9R3 .970, .959 .950 .942 .936 1.667 1.5-3 1.472 1.412 1.367 1.332 1.'104 1.2B1 1.?62 1.233 1.?1~ 1.195 1.1"2 1.171 1.163 1.155 1.149 '} 1.0 1.5 2.0 2.5 1.230 1.394 1.563 1.749 1.091 .978 .e6~ .1I17 .761 .716 .6/10 .6'i0 .~26 .587 .559 .!i37 .5'1 ,"07 1.242 1.127 1.~40 .974 .922 .8112 .1\4~ .8~3 ./l01 .767 .742 .723 .7eA .~96 9 1.409 1.2 4 1.210 1.147 1.099 1.C61 1.031 1.007 .987 .955 .932 .915 .901 .1'90 9 1.51\5 1.471 1;390 1.329 1.283 1.240 '.21 1.196 1.177 1.l47 1.125 1.101l 1.0 9 <; 1.085 10 1.0 1.5 2.0 2.5 1.4~8 1.3~2 1.2~6 ,c'i'+ ;947 .e76 .62'1 .760 .745 .716 .693 .656 .629 .1>09 .5°3 .'i80 .546 .570 .%1 .':~2 1.1qO 1.075 1.02"; .986 .955 .9~O .'109 .A77 .653 .'135 .620 .A'J9 .1\00 .792 .79~ .7H .774 1.351' 1.290 1.24] 1.207 1.17A 1.1r;4 1.13~ 1.105 1.062 1.0~5 1.n~2 1.~41 1.032 1.025 1.018 l.al~ 1.COP, 1.571 1.512 1.46/\ 1.433 1.'105 1.363 1.364 1.336 1.314 1.?96 1.2 0 5 1.?75 1.266 1.259 1.253 1.247 1.243 .496 .f.R7 ./180 .467 .676 .673 1.07~ 1.0~1I 1.250 1.069 .9~6 .666 .795 .736 .693 .656 .62 6 .600 .561 .532 .510 .4"3 .479 .468 1.3~5 1.20" 1.0~0 1.003 .936 .,.,O~ .042 .fl09 .71'2 .759 .72'1 .699 .LAO .6L5 .<;';3 .643 1.514 1.356 1.241 1.1'16 1.093 1.044 1.006 .97"> .9'\0 .929 .890 .874 .1'56 .B~2 .1'31 .822 1.679 1.517 1.403 1.320 1.259 1.213 1.176 1.147 1.1'4 1.104 1.074 1.052 1.035 1.0'1 1,011 1.002 .'+59 .f,35 .1l14 .994 .507 .71A .93C 1.1"~ .479 .;72 .R66 1.062 1.0~7 .451 .h28 .807 .988 .444 .622 .£<02 .91'2 .!Gl .713 .92!i 1.139 .47~ ."1.,7 .6f>6 .861 .'f,1 .oS6 l.e"2 .436 ."'1~ .797 .976 'oJ 0'\ TABLE 5.1.1 (Continued) b = 2.0 ex = 0.05 al f a2 1 II II 9 10 12 111 16 18 20 22 211 26 2R 30 .5/17 1.035 .923 .832 .760 .702 .656 .618 1.1117 1.033 .9115 .876 .1523 .7150 .718 .7116 1.273 1.1~8 1.072 1.007 .957 .060 .~17 .R1l6 0 1.110'1 1.2 11 1.210 1.11&8 1.100 1.063 1.033 1.0n8 .561 .6911 .839 .'10/1 .520 .651l .P05 .'157 .490 .632 .781 .93'1 .467 .612 .7(,3 .'n7 .41&" .5'16 .7'1/1 .903 .'135 .'583 .737 .1113 .561& .719 .A75 .1&05 .557 .712 ./1&'1 .390; .q~2 .'123 .573 .727 .883 .392 .cf.lE• -fO 1 .q'lP. .532 .'190 .'15'1 .581 .710 .6 .. 4 ."~5 .!i)O .691 ./126 .'1'6 .'1 .. 3 .676 .8\2 .'101 .'130 .66'1 ./l01 .389 .519 .655 .792 .379 .511 .6"6 .76'1 .371 .'503 .363 .1&97 .63! .771 .1&11 .3 eU .~19 .502 .6:>0 .7 .. 1 .376 .'189 .... 08 .729 .363 ... 78 .353 .469 .!CJA .S~9 .720 .712 .344 ... 61 .'182 .705 .'55 , .. "i6 .562 .671 .3'12 .661 .331 ."35 .5 .. 3 .J;!:3 .322 ... 27 .536 .646 .326 .'117 .51'1 .613 .314 .407 .505 .605 .305 .399 .'197 .598 .297 .392 ... 91 .290 .!!6 .~9," .~~7 .291 .H6 .282 .36 9 .=4~~ .n5 .362 .4 C,! 2 5 6 12 1.0 1.5 2.0 2.5 1.166 1.292 1.1127 1.568 14 1.0 1.3 2.0 2.5 .807 .735 .676 1.137 1.Cll .8'19 .77/1 .903 .833 1.2"6 1.105 .9°2 .9 .. 4 .892 1.~~2 1.212 1.0'17 1.010 1.~~q, 1.327 1.213 1.128 1.06" 1.015 7 .629 .73'1 .851 .977 8 .~4?1 .... 9'1 ./119 .""6 .5'19 .670 .702 .921 .~ .. 6 .~Oll .770 .qOD .736 .B69 .63C? ,717 .5~1 .71}f, .P"'~ • 357 ."'H .~2e .7&7 16 1.0 1.5 2.0 2.5 .759 .715 1.115 .992 .81'1 1,210 1.013 .960 .670 .600 1.31~ 1.165 1.051 .963 .895 1,"19 1.265 1.150 1.065 1.000 .656 .74'1 .842 .949 .61)9 .699 .801 .910 .570 .663 .767 .A7B .537 .633 .739 .85 2 .510 .6nll .717 .A31 .467 .569 .661 .HIl ."35 .541 .655 ,774 18 1.0 1.5 2,0 2.5 .979 ,866 .774 1.0'111 ./1 .. 6 .9"16 1.11l1 1.C"1I .926 1.271 1.12q 1.C14 1.367 1.21S 1.lnl 1.015 .700 .774 ./157 .949 .6 .. 1 .71'1 .804 .1197 .593 .672 .761 .857 .'153 .1',35 .726 .1124 .521 .604 .6'}8 .7?8 .493 .579 .1',74 .776 .4'19 .638 .7'12 ."1" .510 .611 .717 .3ql ... 117 .'191 .,;98 .371 ,4'10 .'575 .6A3 20 1.0 1.5 2.e 2.5 1.0/1'1 .967 .6'15 1.\51l 1.02/\ .917 1,23':' 1.0'1/1 .9P6 1.32'1 1.176 1.01'.2 .H\O\ .629 .636 .772 .656 .560 .650 .729 .81'1 .5'10 .612 .5n7 .'179 .754 .827 .908 .5~1 .~55 .781 .61'.14 .75'1 .6'10 .731 .401 ."8'1 .575 .670 .376 ."61 .55'1 .651 .3!'>5 ..... 3 .5'8 ,636 .339 .429 .1'.93 .'135 .5114 .603 .696 22 1.0 1.5 2.0 .570 .632 .702 .779 .530 .0;94 .'.66 .7'15 .496 .51;2 .6:'17 .717 .'168 .536 .612 .69'1 .423 ."94 .57'1 .659 .389 ."63 .300 .3~4 .l,d~2 .573 .413 .565 .1~66 .~12 .326 .406 .'193 .564 .3q~ .632 .3 .. 2 ,421 .507 .596 .312 .:'45 .363 ."'10 .5:14 2.~ 1.073 .957 1.13 0 1.012 1.21.~ 1.C7/: 1.283 1.144 .5~9 .s~;:. .~46 24 1.0 1.5 2.0 2.5 1.064 .950 1.12'1 .999 1.190 1.0'55 1.2~0 1.117 ."59 .520 .589 .663 ."13 ."77 .549 .1'.25 .379 ..... 6 .520 .599 .3 .. 3 ."22 ... 98 .3"11 ,'In3 ,'1'1 .389 ,t461 .~~9 .210 .3 .. 9 .'132 .~79 .5"'3 ,550 .289 .365 .'1 .. 7 .'531 .279 .356 ."~1 .300 .37'5 .'156 .539 .'523 .~17 .263 .!t+3 ."26 ."i12 26 1.0 1.5 2.0 2.5 .943 .969 1.~71 1.038 1.235 1.09'1 .451 .506 .40': .3?2 .5hl) ."07 ."77 .'\'\1 .3~7 .5~" .371 .'131 .'199 .572 .~L.;Lt .5":~ .30'5 .H2 ."'10; .'\21 .2 9 1 .359 .U33 .219 .3'19 .42'1 .501 .26<; .3 4 0 ,"16 .49'1 .26 C· .332 .409 .315 .37'1 .440 .510 .:197 • ~~8 .2B2 .42~ • 'I ,260 .325 .396 •• 96 ,4e5 .270 .33'1 .40'1 ,476 ,3011 .3"3 .42'1 .'90 .347 ,'109 ... 74 .275 .263 .322 • ~·e7 .4'3'1 1.0% 1.110 .762 .P26 ./\97 .975 .845 '.752 .&09 .9~1 .873 .?F.2 .343 1.0~0 .678 .737 .875 .618 .679 .747 .621 .670 .723 .782 .8"7 .610 .665 .726 .793 .561 .617 .661 .750 .'121 .!'>7" .'1,.7 .~Ll4 .614 1.0~'I .7 .. 4 .796 .1\53 .916 ,71'1 .6~7 .8"11 .877 .'}27 .9P2 .738 .755 .1137 .!!'13 .663 .711 ,603 .652 .701\ .76'1 .554 .605 .662 .725 .513 .51'.6 .1',25 .F.9n .479 .61'.1 .637 .44'. .495 .t.C'." .473 .522 .578 .837 .81l6 .943 .(\Q2 • '7(>S .1'2'1 28 1.0 1.5 2.0 2.5 1.0"9 1.099 .937 .97'1 1.025 1.213 1.075 .825 .867 .'H3 .964 .732 .775 • 657 .701 .~22 .7~O .87" 30 1.0 1.5 2.0 2.5 1.0'14 .820 1.08~ .B~9 .727 .767 .1\10 ,0';8 1.1~'1 .933 .971 1.1 .. 0 1.013 1.195 1.059 .902 .9"13 .~'.. 7 .548 .'59'1 .E'I7 .70" .'\07 .'i55 .eo'l .597 .6'12 .en • 7'.. 9 .r,6p .6~9 .61'1 .652 .693 .738 .7Po7 .5°1 .633 .680 .731 .5"2 .585 .63'1 .666 .0;01 ,5"6 .'\95 .650 .41'.7 .513 .51'.'1 ,620 .'133 ,"85 .:,52 .53~ ,595 .53~ ... b3 .'129 ."\99 .... 35 .7'15 .398 .451 .511 .576 .363 .'119 ."01 .336 .394 .54~ ,526 .392 .357 .441 .496 ,556 .'IO~ .:130 .3113 .442 .'505 .45~ .527 .4sa .""9: .5'4 .~.~d .~25 .623 .4U4 .~S2 .~lC .~45 ~ .. .333 .3S7 ,463 .336 .33C .~~l.l .41&9 .'176 .699 .~71 .31~ .42~ .30!! ."1'1 .52'1 .6'1e .1,3~ .~94 .~24 .1f'\F: .253 .~26 .4ij3 ."+eA. .4-~2 • .,6S .251 .31P • "0 _r. • ."e2 .2 .... .311 .253 ,313 .378 .24'" .305 .311 .236 ,'I4~ .~4("l • .!.,Q! .U':E.29~ .~65 ~ 34 '-J '-J 78 value k o and a(k ) = Pr(fU 0 l + U > k ) was evaluated. z- 0 Iteration pro- ceeded, producing successive values kl,.••• ,k , until Ik - k _ ' < n n n l .00005. Then k n value k-. n was truncated to three decimals--call the truncated The values a(k-) and a(k+) n n = a(k-n + 0.001) were evaluated. If a(k-) was closer than a(k+) to the desired a-level, k- was chosen n n n as the tabular value, k; otherwise k+ was chosen. n The rounding of k to three decimals produces an error of la(k)-al, which does not exceed 0.00020 if a a = = .01 is the desired level of the test and 0.00060 if .05 is the desired level. For most values the errors are substan- tially smaller, usually less than one-half the stated maximum for the significance level. For values not included in the tables, interpolation is sometimes useful. Interpolation over values of b does not lead to good approximations, but interpolations over the parameters aI' a , and f 2 does result in quite accurate values for k. larger values of f is also feasible. of a l and a 2 Extrapolation for some Extrapolation for larger values was not done since it seems clear from the values in Table 5.1.1 that extrapolation to values of a l and a 2 which are reason- ably close to the maximum tabular value of 30 should yield good approximations to the true value of k. Table 5.1.2 presents values of parameters for which interpolation or extrapolation was performed. (Since the parameters a l and a 2 are interchangeable if f results for certain examples of interpolation over a sponding examples of interpolation over a .) 2 = 1, the apply to corre- l The asterisk indicates the parameter on which interpolation or extrapolation was performed. The value of k, say k, which was found by interpolation or extrapolation " is also given along with the corresponding error, la(k)-al. 79 TABLE. 5 .1. 2 ESTIMATED CRITICAL VALUES FOR TYPE D AND TYPE F TESTS FOR CERTAIN INTERPOLATED AND EXTRAPOLATED VALUES OF THE PARAMETERS a , a , ANDf 2 1 A 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.05 0.05 0.05 0.01 0.01 0.01 0.01 0.05 0.05 0.05 0.01 0.05 0.05 0.05 0.01 0.01 0.01 b a 1.0 2.0 1.0 2.0 1.0 1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0 1.0 1.0 1.0 1.0 2.0 2.0 2.0 2.0 1.0 1.0 2.0 1.0 1.0 1.0 1.5* 1.5* 1. 5* 1.5* 1.5* 29.0* 29.0* 29.0* 29.0* 29.0* 29.0* 29.0* 30.0 30.0 30.0 1.0 1.0 30.0 30.0 30.0 1.0 1.0 30.0 30.0 1.0 1.0 1.0 30.0 30.0 *The 1 a 2 1.0 1.0 1.0 30.0 30.0 1.0 1.0 1.0 1.0 30.0 30.0 30.0 1. 5* 29.0* 29.0* 1.0 30.0 1.0 30.0 30.0 1.0 30.0 30.0 30.0 1.0 1.0 30.0 1.0 30.0 f 1.00 1.00 2.50 1.00 2.50 1.00 1.00 2.50 2.50 1.00 1.00 2.50 2.50 2.50 2.50 1.25* 1. 25* 1.25* 1.25* 1.25* 2.25* 2.25* 2.25* 2.25* 3.00* 3.00* 3.00* 3.00* 3.00* k 1. 770 1.875 3.102 1.063 2.414 1.040 1.109 1.171 1.349 0.209 0.304 0.407 1.124 0.279 0.436 2.092 1.280 1.055 0.234 0.266 2.997 2.256 0.518 0.255 3.450 3.708 3.001 1.209 0.464 A ICl(k)-al 0.00015 0.00008 0.00011 0.00027 0.00054 0.00020 0.00009 0.00005 0.00012 0.00006 0.00003 0.00010 0.00005 0.00041 0.00036 0.00001 0.00004 0.00001 0.00016 0.00042 0.00010 0.00073 0.00000 0.00025 0.00042 0.00040 0.00042 0.00032 0.00024 asterisk signifies the parameter for which the interpolation or extrapolation was performed. 80 Interpolation was done using Newton's forward-difference and backwarddifference formulas (Conte [1965]). difference formulas. Extrapolation was also done using For interpolation on a l and a , five known values 2 of k were used as interpolating points; for interpolation and extrapolation on f, all four known values of k for specified aI' a , b and a 2 were used. Programming for the critical values in Table 5.1.1 was done in PL-l and the PL-l Scientific Subroutine Package program BDTR was used in the computations. BDTR returns the incomplete beta function ratio I (b,a) for given x, b, and a. x As installed on the University of North Carolina Computation Center's IBM 360/75, BDTR returns incorrect values if b = 1, a ~ 1, and x ~ 1. It was possible to avoid using this sub- routine in any such case by substituting the equivalent value of 11 -x (a,b) for I (b,a) as necessary. x 5.2. Numerical Illustration In this section the use of the type D and type F tests in analysis will be demonstrated in an application to data supplied by a pharmaceutical company. Data on the effects of six parallel treatments on diastolic blood pressure measured in sitting position in two hypertensive studies are used. The cell frequencies of the data are as follows: TABLE 5.2.1 DISTRIBUTION OF PATIENTS AMONG SIX HYPERTENSION TREATMENTS IN TWO CLINICS Treatment 4 1 2 5 6 3 1 10 9 7 9 8 9 N i 52 2 11 11 9 9 10 12 62 Clinic 81 The observation x value (Y ijk was obtained as the difference of the pre-treatment ijk ) and the average of post-treatment values collected during four visits (z, 'k). Therefore, the average decrease in diastolic blood ~J pressure of the jth patient at clinic i receiving treatment k is given by (i 1,2; j 1, ••• ,N.; k = 1, ••• ,6). ~ Consider the model X'ijk ~ik where lli is the overall mean at clinic i, effect at clinic i (after Let ~i ~eparameterization) is the k and E, th treatment 'k is the error. ~J be a N x 1 vector of the observations x for clinic i and ijk i = let fi (5.2.1) = lli + ~ik + £ijk' (lli'~i1' ••• '~i5)'. Then (5.2.1) can be rewritten as Xi == A.~. E. _ _J._1 + -1 th th where the m row of A. corresponds to the m row (patient) of X. ~~ ~~ (clinic i) and is 1 1 0 0 0 0 1 -1 1 1 1 1 0 1 0 0 0 -1 0 0 1 0 0 -1 0 0 0 1 0 -1 0 0 0 0 1 -1 if if if if if if patient patient patient patient patient patient m m m m m received received received received received to. received treatment treatment treatment treatment treatment treatment 1 2 3 4 5 6 and ~i is the vector of errors for clinic i. In each clinic the hypothesis tested is H o where ~ = [25x1: C ~ M ""'" _1 _ == Q __ versus HI: !5x 5] and ~ == 1. ~ -0 -C~. ..... 1 M - Assuming that each error vector has an Ni-variate normal distribution with mean zero and Ni x Ni 82 variance-covariance matrix a 2 I, under"H F. 1 = -1 (N i - 6) H.E. /5, i -1-1 = 1,2, the statistic o has the central F distribution with 5 and (N i - 6) degrees of freedom, where ~i Section 1.1, are scalars in this problem. and ~i' defined as in In this example F (5,46) l = 2.3964, PI = 0.05168; F (5,56) = 2.7523, P2 = 0.02716; and 2 the numbers in parentheses indicate the degrees of freedom. The statistics were calculated using the MGLM program, installed at the Triangle Universities Computation Center, and the P-values were calculated by the TSO program CFrDIS, also at TUCC. Since M = 1, the null hypothesis to be tested is, for each clinic, H a . A. = CC = - -1 -1 0, i = 1,2. For this problem the noncentrality parameters are 8 i C,]-l C s/a~ s.)'[C(A~A.)-l = (C_ -1 - ...1-1 ]. - so that , )-1, -1 2 V. _ - [ C(A.A. C] /a . -1 ... -1-1 ... (5.2.2) i For each clinic the maximum likelihood estimate of was substituted for a~. a~, 1 It is not necessary to assume that o~ = o~, as it would be if a single model were fitted to the data. For the 2 0 = 0 2 versus HI: 0 22 I 02 Problem in this example, the test of H: 0 1 21 leads to rejection of H at a = .10. o In order to calculate the statistic flU l + f 2U2 for the type F test it is necessary to find the values of U , U , f , and f 2 • l 2 l Each Ui 83 is a simple transformation of the F-distributed statistics F.. For 1 this example, since b = 5/2 = 2.5, a 1 = 46/2 = 23.0, and a 2 = 56/2 = 28.0, bF/a 1 2.5(2.3964)/23 U1 = 1+bF /a = 1 + 2.5(2.3964)/23 1 1 0.2067 and bF/a 2 2.5(2.7523)/28 = 0.1973. U = = 1 + 2.5(2.7523)/28 2 1+bF/a 2 In order to find the f.' s it is necessary to calculate y. = tr(V.), _1 1 1 i = 1,2, where the V. 's are of the form given in (5.2.2). _1 the matrices [~(~l~i) -1 ~'] -1 = 1,2, , i are provided by MGLM can be used. [~(~I~i) , (A.A ) -1- i -1 -1 • ~'] To determine -1 the 6x 6 matrices (~~~i) which Since ~ = [~5x1: !5x5]' can be found by deleting the first row and column of The resulting 5x5 matrices were inverted to find 18.981 9.000 18.000 9.038 9.000 15.922 9.000 9.000 9.000 18.000 9.019 9.000 8.961 9.000 16.981 22.983 11.983 22.983 11.951 11. 951 20.854 11. 951 11. 951 11.854 20.854 11. 967 11.967 11.902 11.902 21. 935 [C(A'A )-l C,]-l = - _1_1 - and [C(A' A ) -l C' ]-1 = - _2_2 - "2 " " MGLM also provides NiO i , so that O~ and 0 22 can be " N1 O~ " = of = 2405.379 ==> " N2 o~ == 2057.861 ea~i1y 2405.379/52 = 46.257; " > O~ = 2057.861/62 33.191. found: 84 It follows, therefore, that = 87.884/46.Z57 = 1.900 and, similarly, YZ = tr(~Z) ~ 3.30Z. The parameters Y and YZ having been found, it is possible to calculate l = (Z3.0 + Z.5)(1.900)/Z.5 = 19.380; f Z = (a Z + b) YZ/b = (Z8.0 + Z.5)(3.30Z)/2.5 40.284. ~l (al + b) Yl/b = Finally it is possible to calculate the statistic for the type F test. Since f Z > f l , the form fZUZ/fl + Ul = k is used. It follows from the above that k = 40.Z84 = Since k ~ (0.1973)/19.380 + 0.Z067 Z.079 (0.1973) + 0.Z067 = 0.617. 1.0 and kfl/f Z ~ 1.0, a program using the algorithm to provide significance levels for Case 3 situations can be used to determine that, for this problem, Pr(Z.079 U + U z l ~ 0.617) = .00869. Therefore, the overall test is significant at a = .01 whereas neither of the individual F statistics, F and FZ' is significant at that level. l By contrast, Fisher's statistic for this problem is -2 log (PlP Z) = -21og (0.05168 x 0.OZ716) = 13.1374. This statistic is distributed as a chi-squared with four degrees of freedom under H. o By 85 use of CFTDIS it can be found that Pr~x: ~ 13.1374) = .01062. For this problem, the type F test is found to be more sensitive than Fisher's test to deviations from the null hypothesis, even though there are only two statistics combined. This result is of particular interest in light of the results of Section 3.3, where it was shown that the type D and type F tests are more efficient assuming that there are a large number of statistics to be combined. The above application illustrates the method of calculating a statistic for the type F test. Since the type D test coincides with the type F test if indeed the V. are proportional, the above method of -~ calculation, with y. ~ = tr(V.),can be used for type D tests also. -~ CHAPTER VI SUMMARY AND SOME SUGGESTIONS FOR FURTHER RESEARCH In this dissertation two locally most powerful procedures were found for combining independent statistics which arise in the analysis of certain multiparametric and multivariate problems. The type D test for combining beta-distributed statistics was found for the case of a balanced design; the type F test was introduced for use in the case of an unbalanced design. The overall statistic for type D and type F tests is, in each case, a linear combination of the beta-distributed statistics to be combined. The exact distribution of the statistic was found in the two-sample case; the accuracy of using normal approximations for the general s-sample case was also considered. It was shown that for local alternatives, the type D test is more efficient than Fisher's test for combining statistics, according to Bahadur asymptotic relative efficiency, if there are a large number of smallsample statistics to be combined. Under the same conditions, it was shown that the type F test is more efficient than Fisher's test according to ciency. a~ average measure of Bahadur asymptotic relative effi- In a numerical example, the theor~tical developments were applied to some phannaceutical data. Several areas of future research may be suggested. It would be useful to work out the details of the exact distribution of the type D and type F statistics for larger values of s. As noted 87 earlier, however, this task would be tedious and the calculations necessary to find critical values would. be expensive. It would also be useful to develop the details of the distributions of the type D and type F statistic if the statistics to be combined have chi-squared distributions. If s = 2, the exact distribution should be reasonably simple to calculate, with more complicated results for larger values of s. Another area of further research is consideration of the extensions mentioned in Section 3.4. The exact distributions of the combined test statistics mentioned there may be sufficiently complicated to warrant study of approximations. As mentioned earlier, Littell and Folks have shown that if there are a finite number of large-sample statistics to be combined, Fisher's method is more efficient than other methods according to Bahadur asymptotic relative efficiency; the locally most powerful combination procedures were shown to be more efficient if there are a large number of finite-sample statistics to be combined, for local alternatives. Research into the relative efficiencies of Fisher's test and the locally most powerful tests for combining an intermediate number of medium-sized sample statistics for local and non-local alternatives would be illuminating. APPENDIX A USEFUL MATHEMATICAL RESULTS Lemma A.l. If U. has a beta distribution with density 1 e. Su.1 specified 1 in (1.1.2), then (2.3.4) and (2.3.5) are equivalent. s Proof. In (2.3.4) II h=l form of (1.1. 2). eh Su is the product of s infinite series of the h Since each of these series is a positive term con- vergent series, each is absolutely convergent and (Taylor and Mann [1972, p. 635]) the order of summation can be interchanged. It remains to be shown that each of the infinite sums can be taken outside the integral. If b ~ 1 and a j ~ 1, then the density of each U is uniformly j If b = 1/2 then for every j, bounded and the interchange is immediate. the density of U. goes to infinity as U. J J the density goes to infinity as U t 1. j dU j ~ O. J However, since for all j = 1, the dominated convergence theorem (Cramer [1945, p. 45]) applied. Therefore, each summation can be taken outside the integral and the lemma holds. Lemma A.2. If a. = 1/2 for some j, o -< are as defined in (1.1.3), then u.1<- 1, i 89 u ! eO eO ,w 1 u u 2 dUldu Z = r (1) u 2 u = (2) 2 r U(2)(u ) 1 Z r ui 1 ) (u ) eOu eOu 1 Z (2) l u(2)(u) Z 1 r U~l)(u1) (1) u l du du l 2 Z eOu 13u0 1 dU du , 2 l 2 where the limits of integration depend on the boundary w Proof. As long as the function w. is integrable in w, the result holds (Taylor and Mann [1972, p. euo , c The beta densities, 13 0 u and their product, are integrable when w c w. and l 2 Lemma A.3. 00· Let f(x) = Ln=O anx convergence R > O. n be a power series with radius of The function f is continuous in the open interval of convergence of the series. If p and q are points in [O,R), p < q, then r p q ~oo f(x)dx = un=O an n+l 0+1 .;Iq'--_ _ -~P-_ n+l .• If q = R, the result holds if the right hand side is convergent. Reference. Discussion. Taylor and Mann [1972, p. 665]. In all the instances in which this lemma is used in the dissertation, the series is a Taylor series in which a n = fen) (O)/n!, where fen) denotes the nth derivative of f. 90 Lemma A.4. If 0 ~ ~ x c/d and if n > 0, then where '* n {n, if n is an integer; ~,otherwise. = Reference. Taylor and Mann [1972, p. 632]. Lemma A.5. Let 0 1, b > 0, and a > O. If S (b,a) = f x u b-1 (l-u) a-I du ~ ~ x x 0 is the incomplete beta function and I (b,a) x Sx (b,a)/S(b,a) = is the incomplete beta function ratio, then (a-I) '* Sx (b,a) = L S(b,a) =L i=O 1 = (a-I) '* a-I i=O 1. ~ x b-1 x 0 (l-u) i=O a-I (A.5.1) ,and du (A. 5.2) 1. (a-I) '* a-I (_l)i b+i Q(b,a) L ( i ) b+i x , By Lemma A.3 and A.4, S (b,a) = f u (_l)i ( " ) b+" where (a-I) '* is defined in Lemma A.4. Proof. b+1." x 1. 1 I (b,a) x (-1) i (a~ ) b+1." (A.5.3) 91 (a-I) * =L i=O i ( a-I) H2.. x b+i i b+i This verifies the first claim. the second claim follows by setting x = S(b,a)., so = 1 in the first result. third claim follows by definition of I (b,a). x Lemma A.6. f\ (b,a) By definition, The The completes the proof. If p, q, a, and b are real numbers such that q > p, b > 0, and a > 0, then f q (u_p)b-l (q_u)a-l du (q-p) a+b-l S(b,a). p Discussion. and q = 1, Although the standard beta distribution specifies p =0 a family of beta distributions is given by density functions of the form (Johnson and Kotz, Vol. 2 [1970, p. 37]). a standard Lemma A.7. If V = (U-p)/(q-p), V has beta distribution. If x ~ 1, b > 0, a > 0, and c > 1, then CA.7.1) converges. Proof. vergent. If a is an integer the series is finite and, therefore, cona-I If a is not an integer, then notice that ( i )(-1) positive or negative for all i > (a-I). i is either If [al is even, the terms are positive; if [a] is odd, the terms are negative, where [a] is the greatest integer in a. If [a] is odd, multiply the .series by (-1), 92 which will not affect the convergence or divergence of the series. Therefore, without loss of generality, .assume that the series is a positive term series in the tail. It will be shown that (A.7.l) converges for non-integral a by showing that 00 I~ (a-I) (_l)i xb+i Q (b+i) ~=O i ~l/x'c I< where M and M are positive constants. l 2 00 + IL i=[a] 00 M M I~ (a-I) (-1) 1 + 2 ui=[a] i b+i i I (A 7 2) •• By Lemma A.5, . I (a-I i )(-1) i x b+i f3 l / x (b+1,c). (A. 7 • 3) Next, by Lemma A.5, 'Ie (c-l) (l)j] • + L (c-l) (-l)j b+i b+i+j x j=[e] j (A. 7 .4) 93 [c-l] . E (c-l) (l)J, j=O j x The first series in (A.7.4) is finite and bounded by M 3 which is finite. If c is an integer, the second series in (A.7.4) is O. 00 Otherwise, recall that Mann [1972, p. 632]). E (c~l)(_l)j converges if c-l > 0 (Taylor and J j=O 00 \"'L. (c-l) (-l)j converges. Since Therefore, j= [c] j all terms in this series are of the same sign, 00 00 j b+i 1· IE (c- )(-1) b+i+j (;z)JI < IE j=[c] j j=[c] 1 (Cj 1 )(_l)J. I = 4 < M 00 • (A.7.5) Let M = M + M , which is finite. 2 3 4 Then Let M be the first term on the right hand side of (A.7.3). l M < l Clearly Then (A.7.2) follows from (A.7.3) in light of (A.7.6). Since 00 a-I (_l)i , ~, M and Ei=[a]( i )b+i are all finite, the lemma holds as long as 2 c > 1. (Note that in all instances when the lemma is used in the text, 00. c > 1.) Lemma A.8. If a > 0, b > 0, c > 0 and x ~ 1, (a-I) * a-I (_l)i b+i i:O ( i )b+i x f3 l / x (b+i+1, c) converges. Proof. The proof proceeds on the principles of the proof of the previous lemma. However, (A.7.4) is replaced by 94 x b+i Sl/x(b+i+l,c) < IiC-l] (c-l) (-l)j J. b+i+l+J· j=O (!.)j+ll x (c-l) * c-l (-l)j + Ir ( j )b+i+l+j j=[c] I· (A.8.l) If c is not an integer, the second term on the right-hand side is the tail of the series IX) c-l (-l)j r (J. )b+i+l+j = S(b+i+l,c),. which is a decreasing j=O function of i and, therefore, is bounded. (A.8.l) is set equal to M < 2 i IX) Ir i=O 00, If the right-hand side of it is then clear that b+i xb+~ Sl/x(b+i+l,c) I < (a-il) (-1) ~ i 00 M 1 + M Ir 2 i=[a] (a-:-l) (-:) ~ b+~ I, (A. 8.2) where M is the absolute value of the first a terms of the series on l the left-hand side. Since the series on the right-hand side is the tail of the convergent series (A.S.2), (A.8.2) is finite and the lemma holds. Lemma A.9. If I (b,a) is an incomplete beta function ratio defined in x Lemma A.S, then I x (b,a) = 1 - 11-x (a,b'. { Reference. K. Pearson [1968, p. vii]. BIBLIOGRAPHY Anderson, T. W. Analysis. [1958]. An Introduction to Multivariate Statistical Wiley, New York. Bahadur, R. R. [1960]. Stochastic comparison of tests. Statist. 31, 276-295. Ann. Math. Bahadur, R. R. [1967]. Rates of convergence of estimates and test statistics. Ann. Math Statist. 38, 303-324. Bhattacharya, N. [1961]. Sampling experiments on the combination of independent x2-tests. Sankhya A 23, 191-196. Birnbaum, A. [1954]. Combining independent tests of significance. J. Amer. Statist. Ass. 49, 559-574. Birnbaum, A. [1955]. Characteristizations of complete classes of tests of some multiparametric hypotheses, with applications to likelihood ratio tests. Ann. Math. Statist. 26, 21-36. Conte, S. D. [1965]. New York. Elementary Numerical Analysis. McGraw-Hill, Cramer, H. [1946]. Mathematical MethodS of Statistics. University Press, Princeton. Princeton Davies, R. B. The weighted combination of independent normal test statistics. (Unpublished). Fisher, R. A. [1932]. Statistical Methods for Research Workers (4th and Later Editions). Oliver and Boyd, London. Fisz, M. [1963]. Probability Theory and Mathematical Statistics. Wiley, New York. Good, I. J. [1955]. On the weighted combination of significance tests. J. R. Statist. Soc. B 17, 264-265. [1969]. Introduction to Matrices with Applications in Statistics. Wadsworth, Belmont, California. Graybill, F. A. Gupta, A. K. [1971]. Noncentral distribution of Wilks' statistic in MANOVA. Ann. Math. Statist. 42, 1254-1261. 96 Hill, A. B. [1971]. Principles of Medical Statistics (9th Edn.). Oxford University Press, New York. Isaacson, S. L. [1951]. On the theory of unbiased tests of simple statistical hypotheses specifying two or more parameters. Ann. Math. Statist. 22~ 217-234. Johnson, N. L. and Kotz, S. [1970]. Continuous Univariate Distributions~ Vols. 1 and 2. Houghton Mifflin, Boston. Kincaid, W. M. [1962]. The combination of test statistics based on discrete distributions. J. Amer. Statist. Ass. 57~ 10-19. Lancaster, H. O. [1949]. The combination of probabilities arlslng from data in discrete distributions. Biometrika 36~ 370-382. Lancaster, H. O. [1961]. The combination of probabilities: an application of orthonormal functions. Austral. J. Statist. 20-33. Liptak, T. [1958]. On the combination of independent tests. 3~ 171-197. 3~ Magyar Tud. Akad. Mat. KutatO Int. Kozl. Littell, R. C. and Folks, J. L. [1971]. Asymptotic optimality of Fisher's method of combining independent tests. J. Amer. Statist. Ass. 66~ 802-806. Littell, R. C. and Folks, J. L. [1973]. Asymptotic optimality of Fisher's method of combining independent tests II. J. Amer. Statist. Ass. 68~ 193-194. Mantel, N. [1963]. Chi-squared tests with one degree of freedom; extensions of the Mantel-Haenszel procedures. J. Amer. Statist. Ass. 58, 690-700. Morrison, D. F. [1967]. Hill, New York. Multivariate Statistical Methods~ McGraw- Naik, U. D. [1969]. The equal probability test and its applications to some simultaneous inference problems. J. Amer. Statist. Ass. 64~ 986-998. Neyman, J. and Pearson. E. S. [1936]. Contributions to the theory of testing statistical hypotheses, part I. In Statistical Research Memoirs~ Vol. 1. Issued by the Department of Statistics, University of London, University College, pp. 1-37. Neyman, J. and Pearson, E. S. [1938]. Contributions to the theory of testing statistical hypotheses, parts II and III. In Statistical Research Memoirs~ Vol. 2. Issued by the Department of Statistics, University of London, University College, pp. 25-57. 97 Oosterhoff, J. [1969]. Combination of One-Sided Statistical Tests. Tract No. 28, Amsterdam Mathematical Centre, Amsterdam. Pearson, E. S. [1938]. The probability integral transformation for testing goodness of fit and combining independent tests of significance. Biometrika 30~ 134-148. Pearson, E. S. [1950]. On questions raised by the combination of tests based on discontinuous distributions. Biometrika 37~ 383-398. Pearson, K. [1933]. On a method of determining whether a sample of given size n supposed to have been drawn from a parent population having a known probability integral has probably been drawn at random. Biometrika 25~ 379-410. Pearson, K. (Ed.) [1968]. Tables of the Incomplete Beta Function (2nd Edn.). Cambridge University Press, London. (1st Edn., 1934). Pi11ai, K. C. S. and Young, D. L. [1971]. On the exact distribution of Hote11ing's generalized T~. J. Multivariate Anal. 1~ 90107. [1971]. Analysis and Design of Certain Quantitative Multiresponse Experiments. Roy, S. N., Gnanadesikan, R. and Srivastava, J. N. Pergamon Press, New York. Schaafsma, W. [1968]. A comparison of the most stringent and the most stringent somewhere most powerful test for certain problems with restrictive alternative. Ann. Math. Statist. 39, 531-546. Sen, P. K. [1968]. Robustness of some nonparametric procedures in linear models. Ann. Math. Statist. 39~ 1913-1922. Siotani, M. [1971]. An asymptotic expansion of the non-null distribution of Hote11ing's generalized T~-statistic. Ann. Math. Statist. 42~ 560-571. Sugiura, N. and Fujikoshi, Y. [1969]. Asymptotic expansions of the non-null distributions of the likelihood ratio criteria for multivariate linear hypothesis and independence. Ann. Math. Statist. 40, 942-952. Taylor, A. and Mann, W. R. [1972]. Advanced Calculus. Publishing, Lexington, Massachusetts. Tippett, L. H. C. [1931]. Norgate, London. The Methods of Statistics. Xerox College Williams and 98 Van E1teren, Ph. [1958]. On the combination of independent two sample tests of Wilcoxon. BuZZ. Intern. Statist. Inst. 37, 351-361. Van Zwet, W. R. and Oosterhoff, J. [1967]. On the combination of independent test statistics. Ann. Math. Statist. 38, 659-680. Wa1d, A. [1943]. Tests of statistical hypotheses concerning several parameters when the number of observations is large. Trans. Amer. Math. Soc. 54, 426-482. Wallis, W. A. [1942]. Compounding probabilities from independent significance tests. Econometrica 10, 229-249. Wey1, H. [1939]. 461-472. On the volume of tubes. Amer. J. of Math. 61, Wi1k, M. B. and Shapiro, S. S. [1968]. The joint assessment of normality of several independent samples. Technometrias 10, 825-839. Wilkinson, B. [1951]. A statistical consideration in psychological research. Psych. BuZZ. 48, 156-158. Zelen, M. and Joel, L. S. [1959]. The weighted compounding of independent significance tests. Ann. Math. Statist. 30, 885-895. two
© Copyright 2024 Paperzz